Fine-tune Llama 3 for text generation on Amazon SageMaker JumpStart

Published:


Generative artificial intelligence (AI) models have become increasingly popular and powerful, enabling a wide range of applications such as text generation, summarization, question answering, and code generation. However, despite their impressive capabilities, these models often struggle with domain-specific tasks or use cases due to their general training data. To address this challenge, fine-tuning these models on specific data is crucial for achieving optimal performance in specialized domains.

In this post, we demonstrate how to fine-tune the recently released Llama 3 models from Meta, specifically the llama-3-8b and llama-3-70b variants, using Amazon SageMaker JumpStart. The fine-tuning process is based on the scripts provided in the llama-recipes repo from Meta, utilizing techniques like PyTorch FSDP, PEFT/LoRA, and Int8 quantization for efficient fine-tuning of these large models on domain-specific datasets.

By fine-tuning the Meta Llama 3 models with SageMaker JumpStart, you can harness their improved reasoning, code generation, and instruction following capabilities tailored to your specific use cases.

Meta Llama 3 overview

Meta Llama 3 comes in two parameter sizes—8B and 70B with 8,000 context length—that can support a broad range of use cases with improvements in reasoning, code generation, and instruction following. Meta Llama 3 uses a decoder-only transformer architecture and new tokenizer that provides improved model performance with 128,000 context size. In addition, Meta improved post-training procedures that substantially reduced false refusal rates, improved alignment, and increased diversity in model responses. You can now derive the combined advantages of Meta Llama 3 performance and MLOps controls with Amazon SageMaker features such as Amazon SageMaker Pipelines and Amazon SageMaker Debugger. In addition, the model will be deployed in an AWS secure environment under your virtual private cloud (VPC) controls, helping provide data security.

SageMaker JumpStart

SageMaker JumpStart is a powerful feature within the SageMaker machine learning (ML) environment that provides ML practitioners a comprehensive hub of publicly available and proprietary foundation models (FMs). With this managed service, ML practitioners get access to a growing list of cutting-edge models from leading model hubs and providers that they can deploy to dedicated SageMaker instances within a network isolated environment, and customize models using SageMaker for model training and deployment.

Prerequisites

To try out this solution using SageMaker JumpStart, you’ll need the following prerequisites:

Fine-tune Meta Llama 3 models

In this section, we discuss the steps to fine-tune Meta Llama 3 models. We’ll cover two approaches: using the SageMaker Studio UI for a no-code solution, and utilizing the SageMaker Python SDK.

No-code fine-tuning through the SageMaker Studio UI

SageMaker JumpStart provides access to publicly available and proprietary foundation models from third-party and proprietary providers. Data scientists and developers can quickly prototype and experiment with various ML use cases, accelerating the development and deployment of ML applications. It helps reduce the time and effort required to build ML models from scratch, allowing teams to focus on fine-tuning and customizing the models for their specific use cases. These models are released under different licenses designated by their respective sources. It’s essential to review and adhere to the applicable license terms before downloading or using these models to make sure they’re suitable for your intended use case.

You can access the Meta Llama 3 FMs through SageMaker JumpStart in the SageMaker Studio UI and the SageMaker Python SDK. In this section, we cover how to discover these models in SageMaker Studio.

SageMaker Studio is an IDE that offers a web-based visual interface for performing the ML development steps, from data preparation to model building, training, and deployment. For instructions on getting started and setting up SageMaker Studio, refer to Amazon SageMaker Studio.

When you’re in SageMaker Studio, you can access SageMaker JumpStart by choosing JumpStart in the navigation pane.

In the JumpStart view, you’re presented with the list of public models offered by SageMaker. You can explore other models from other providers in this view. To start using the Meta Llama 3 models, under Providers, choose Meta.

Public Models

You’re presented with a list of the models available. Choose the Meta-Llama-3-8B-Instruct model.

Meta Model Provider

Here you can view the model details, as well as train, deploy, optimize, and evaluate the model. For this demonstration, we choose Train.

Meta Llama 3 8B Instruct Details

On this page, you can point to the Amazon Simple Storage Service (Amazon S3) bucket containing the training and validation datasets for fine-tuning. In addition, you can configure deployment configuration, hyperparameters, and security settings for fine-tuning. Choose Submit to start the training job on a SageMaker ML instance.

Fine-tune model

Deploy the model

After the model is fine-tuned, you can deploy it using the model page on SageMaker JumpStart. The option to deploy the fine-tuned model will appear when fine-tuning is finished, as shown in the following screenshot.

Finetuning Finished Screen

You can also deploy the model from this view. You can configure endpoint settings such as the instance type, number of instances, and endpoint name. You will need to accept the End User License Agreement (EULA) before you can deploy the model.

Deploy Model Screen

Fine-tune using the SageMaker Python SDK

You can also fine-tune Meta Llama 3 models using the SageMaker Python SDK. A sample notebook with the full instructions can be found on GitHub. The following code example demonstrates how to fine-tune the Meta Llama 3 8B model:

import os
import boto3
from sagemaker.session import Session
from sagemaker.jumpstart.estimator import JumpStartEstimator

# To fine-tune the Llama 3 70B model available on JumpStart, please change model_id to `meta-textgeneration-llama-3-70b`.
model_id = "meta-textgeneration-llama-3-8b"
accept_eula = "true"
estimator = JumpStartEstimator(
    model_id=model_id, environment={"accept_eula": accept_eula}
)

# By default, instruction tuning is set to false. Thus, to use instruction tuning dataset you use instruction_tuned="True"
estimator.set_hyperparameters(instruction_tuned="True", epoch="5")
estimator.fit({"training": train_data_location})

The code sets up a SageMaker JumpStart estimator for fine-tuning the Meta Llama 3 large language model (LLM) on a custom training dataset. It configures the estimator with the desired model ID, accepts the EULA, enables instruction tuning by setting instruction_tuned="True", sets the number of training epochs, and initiates the fine-tuning process.

When the fine-tuning job is complete, you can deploy the fine-tuned model directly from the estimator, as shown in the following code. As part of the deploy settings, you can define the instance type you want to deploy the model on. For the full list of deployment parameters, refer to the deploy parameters in the SageMaker SDK documentation.

# for Llama 3 70B models, you can deploy to ml.g5.12xlarge instance type or it will default to ml.p4d.24xlarge
finetuned_predictor = estimator.deploy(instance_type="ml.g5.12xlarge")

After the endpoint is up and running, you can perform an inference request against it using the predictor object as follows:

prompt = "Your prompt goes here"
payload = {
        "inputs": prompt,
        "parameters": {"max_new_tokens": 256},
    }
response = finetuned_predictor.predict(payload)
response.get('generated_text')

For the full list of predictor parameters, refer to the predictor object in the SageMaker SDK documentation.

Fine-tuning technique

Language models such as Meta Llama are more than 10 GB or even 100 GB in size. Fine-tuning such large models requires instances with significantly higher CUDA memory. Furthermore, training these models can be very slow due to their size. Therefore, for efficient fine-tuning, we use the following optimizations:

  • Low-Rank Adaptation (LoRA) – This is a type of parameter efficient fine-tuning (PEFT) for efficient fine-tuning of large models. In this, we freeze the whole model and only add a small set of adjustable parameters or layers into the model. For instance, instead of training all 8 billion parameters for Llama 3 8B, we can fine-tune less than 1% of the parameters. This helps significantly reduce the memory requirement because we only need to store gradients, optimizer states, and other training-related information for only 1% of the parameters. Furthermore, this helps reduce both training time and cost. For more details on this method, refer to LoRA: Low-Rank Adaptation of Large Language Models.
  • Int8 quantization – Even with optimizations such as LoRA, models like Meta Llama 70B require significant computational resources for training. To reduce the memory footprint during training, we can employ Int8 quantization. Quantization typically reduces the precision of the floating-point data types. Although this decreases the memory required to store model weights, it can potentially degrade the performance due to loss of information. However, Int8 quantization utilizes only a quarter of the precision compared to full-precision training, but it doesn’t incur significant degradation in performance. Instead of simply dropping bits, Int8 quantization rounds the data from one type to another, preserving the essential information while optimizing memory usage. To learn about Int8 quantization, refer to int8(): 8-bit Matrix Multiplication for Transformers at Scale.
  • Fully Sharded Data Parallel (FSDP) – This is a type of data parallel training algorithm that shards the model’s parameters across data parallel workers and can optionally offload part of the training computation to the CPUs. Although the parameters are sharded across different GPUs, computation of each microbatch is local to the GPU worker. It shards parameters more uniformly and achieves optimized performance through communication and computation overlapping during training.

The following table compares different methods with the two Meta Llama 3 models.

Default Instance Type Supported Instance Types with Default configuration Default Setting LORA + FSDP LORA + No FSDP Int8 Quantization + LORA + No FSDP
Llama 3 8B ml.g5.12xlarge ml.g5.12xlarge, ml.g5.24xlarge, ml.g5.48xlarge LORA + FSDP Yes Yes Yes
Llama 3 70B ml.g5.48xlarge ml.g5.48xlarge INT8 + LORA + NO FSDP No No Yes

Fine-tuning of Meta Llama models is based on scripts provided by the GitHub repo.

Training dataset format

SageMaker JumpStart currently support datasets in both domain adaptation format and instruction tuning format. In this section, we specify an example dataset in both formats. For more details, refer to the Dataset formatting section in the appendix.

Domain adaptation format

The Meta Llama 3 text generation model can be fine-tuned on domain-specific datasets, enabling it to generate relevant text and tackle various natural language processing (NLP) tasks within a particular domain using few-shot prompting. This fine-tuning process involves providing the model with a dataset specific to the target domain. The dataset can be in various formats, such as CSV, JSON, or TXT files. For example, if you want to fine-tune the model for the domain of financial reports and filings, you could provide it with a text file containing SEC filings from a company like Amazon. The following is an excerpt from such a filing:

This report includes estimates, projections, statements relating to our
business plans, objectives, and expected operating results that are “forward-
looking statements” within the meaning of the Private Securities Litigation
Reform Act of 1995, Section 27A of the Securities Act of 1933, and Section 21E
of the Securities Exchange Act of 1934. Forward-looking statements may appear
throughout this report, including the following sections: “Business” (Part I,
Item 1 of this Form 10-K), “Risk Factors” (Part I, Item 1A of this Form 10-K),
and “Management’s Discussion and Analysis of Financial Condition and Results
of Operations” (Part II, Item 7 of this Form 10-K). These forward-looking
statements generally are identified by the words “believe,” “project,”
“expect,” “anticipate,” “estimate,” “intend,” “strategy,” “future,”
“opportunity,” “plan,” “may,” “should,” “will,” “would,” “will be,” “will
continue,” “will likely result,” and similar expressions.

Instruction tuning format

In instruction fine-tuning, the model is fine-tuned for a set of NLP tasks described using instructions. This helps improve the model’s performance for unseen tasks with zero-shot prompts. In instruction tuning dataset format, you specify the template.json file describing the input and the output formats and the train.jsonl file with the training data item in each line.

The template.json file always has the following JSON format:

{
  "prompt": "<<Prompt goes here along with question or context or instruction>>",
  "completion": "<<completion goes here depending on the activity, for ex: answer for Q&A or summary for Summarization task>>"
}

For instance, the following table shows the template.json and train.jsonl files for the Dolly and Dialogsum datasets.

Dataset Use Case template.json train.jsonl
Dolly Question Answering {
“prompt”: “Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{context}\n\n”,
“completion”: ” {response}”
}
{
“instruction”: “Who painted the Two Monkeys”,
“context”: “Two Monkeys or Two Chained Monkeys is a 1562 painting by Dutch and Flemish Renaissance artist Pieter Bruegel the Elder. The work is now in the Gemäldegalerie (Painting Gallery) of the Berlin State Museums.”,
“response”: “The two Monkeys or Two Chained Monkeys is a 1562 painting by Dutch and Flemish Renaissance artist Pieter Bruegel the Elder. The work is now in the Gemaeldegalerie (Painting Gallery) of the Berlin State Museums.”
}
Dialogsum Text Summarization {
“prompt”: “Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n{dialogue}\n\n”,
“completion”: ” {summary}”
}
{
“dialogue”: “#Person1#: Where do these flower vases come from? \n#Person2#: They are made a town nearby. The flower vases are made of porcelain and covered with tiny bamboo sticks. \n#Person1#: Are they breakable? \n#Person2#: No. They are not only ornmamental, but also useful. \n#Person1#: No wonder it’s so expensive. “,
“summary”: “#Person2# explains the flower vases’ materials and advantages and #Person1# understands why they’re expensive.”
}

Supported hyperparameters for training

The fine-tuning process for Meta Llama 3 models allows you to customize various hyperparameters, each of which can influence factors such as memory consumption, training speed, and the performance of the fine-tuned model. At the time of writing this post, the following are the default hyperparameter values. For the most up-to-date information, refer to the SageMaker Studio console, because these values may be subject to change.

  • epoch – The number of passes that the fine-tuning algorithm takes through the training dataset. Must be an integer greater than 1. Default is 5.
  • learning_rate – The rate at which the model weights are updated after working through each batch of training examples. Must be a positive float greater than 0. Default is 0.0001.
  • lora_r – Lora R dimension. Must be a positive integer. Default is 8.
  • lora_alpha – Lora Alpha. Must be a positive integer. Default is 32.
  • target_modules – Target modules for LoRA fine-tuning. You can specify a subset of [‘q_proj’,’v_proj’,’k_proj’,’o_proj’,’gate_proj’,’up_proj’,’down_proj’] modules as a string separated by a comma without any spaces. Default is q_proj,v_proj.
  • lora_dropout – Lora Dropout. Must be a positive float between 0 and 1. Default is 0.05.
  • instruction_tuned – Whether to instruction-train the model or not. At most one of instruction_tuned and chat_dataset can be True. Must be True or False. Default is False.
  • chat_dataset – If True, dataset is assumed to be in chat format. At most one of instruction_tuned and chat_dataset can be True. Default is False.
  • add_input_output_demarcation_key – For an instruction tuned dataset, if this is True, a demarcation key ("### Response:\n") is added between the prompt and completion before training. Default is True.
  • per_device_train_batch_size – The batch size per GPU core/CPU for training. Default is 1.
  • per_device_eval_batch_size – The batch size per GPU core/CPU for evaluation. Default is 1.
  • max_train_samples – For debugging purposes or quicker training, truncate the number of training examples to this value. Value -1 means using all of the training samples. Must be a positive integer or -1. Default is -1.
  • max_val_samples – For debugging purposes or quicker training, truncate the number of validation examples to this value. Value -1 means using all of the validation samples. Must be a positive integer or -1. Default is -1.
  • seed – Random seed that will be set at the beginning of training. Default is 10.
  • max_input_length – Maximum total input sequence length after tokenization. Sequences longer than this will be truncated. If -1, max_input_length is set to the minimum of 1024 and the maximum model length defined by the tokenizer. If set to a positive value, max_input_length is set to the minimum of the provided value and the model_max_length defined by the tokenizer. Must be a positive integer or -1. Default is -1.
  • validation_split_ratio – If validation channel is None, ratio of train-validation split from the train data must be between 0–1. Default is 0.2.
  • train_data_split_seed – If validation data is not present, this fixes the random splitting of the input training data to training and validation data used by the algorithm. Must be an integer. Default is 0.
  • preprocessing_num_workers – The number of processes to use for preprocessing. If None, the main process is used for preprocessing. Default is None.
  • int8_quantization – If True, the model is loaded with 8-bit precision for training. Default for 8B is False. Default for 70B is True.
  • enable_fsdp – If True, training uses FSDP. Default for 8B is True. Default for 70B is False.

Instance types and compatible hyperparameters

The memory requirement during fine-tuning may vary based on several factors:

  • Model type – The 8B model has the smallest GPU memory requirement and the 70B model has a largest memory requirement
  • Max input length – A higher value of input length leads to processing more tokens at a time and as such requires more CUDA memory
  • Batch size – A larger batch size requires larger CUDA memory and therefore requires larger instance types
  • Int8 quantization – If using Int8 quantization, the model is loaded into low precision mode and therefore requires less CUDA memory

To help you get started, we provide a set of combinations of different instance types, hyperparameters, and model types that can be successfully fine-tuned. You can select a configuration as per your requirements and availability of instance types. We fine-tune all three models on a variety of settings with three epochs on a subset of the Dolly dataset with summarization examples.

8B model

Instance Type Max Input Length Per Device Batch Size Int8 Quantization Enable FSDP Time Taken (Minutes)
ml.g4dn.12xlarge 1024 2 TRUE FALSE 202
ml.g4dn.12xlarge 2048 2 TRUE FALSE 192
ml.g4dn.12xlarge 1024 2 FALSE TRUE 98
ml.g4dn.12xlarge 1024 4 TRUE FALSE 200
ml.g5.12xlarge 2048 2 TRUE FALSE 73
ml.g5.12xlarge 1024 2 TRUE FALSE 88
ml.g5.12xlarge 2048 2 FALSE TRUE 24
ml.g5.12xlarge 1024 2 FALSE TRUE 35
ml.g5.12xlarge 2048 4 TRUE FALSE 72
ml.g5.12xlarge 1024 4 TRUE FALSE 83
ml.g5.12xlarge 1024 4 FALSE TRUE 25
ml.g5.12xlarge 1024 8 TRUE FALSE 83
ml.g5.24xlarge 2048 2 TRUE FALSE 73
ml.g5.24xlarge 1024 2 TRUE FALSE 86
ml.g5.24xlarge 2048 2 FALSE TRUE 24
ml.g5.24xlarge 1024 2 FALSE TRUE 35
ml.g5.24xlarge 2048 4 TRUE FALSE 72
ml.g5.24xlarge 1024 4 TRUE FALSE 83
ml.g5.24xlarge 1024 4 FALSE TRUE 25
ml.g5.24xlarge 1024 8 TRUE FALSE 82
ml.g5.48xlarge 2048 2 TRUE FALSE 73
ml.g5.48xlarge 1024 2 TRUE FALSE 87
ml.g5.48xlarge 2048 2 FALSE TRUE 27
ml.g5.48xlarge 1024 2 FALSE TRUE 48
ml.g5.48xlarge 2048 4 TRUE FALSE 71
ml.g5.48xlarge 1024 4 TRUE FALSE 82
ml.g5.48xlarge 1024 4 FALSE TRUE 32
ml.g5.48xlarge 1024 8 TRUE FALSE 81
ml.p3dn.24xlarge 2048 2 TRUE FALSE 104
ml.p3dn.24xlarge 1024 2 TRUE FALSE 114

70B model

Instance Type Max Input Length Per Device Batch Size Int8 Quantization Enable FSDP Time Taken (Minutes)
ml.g5.48xlarge 1024 1 TRUE FALSE 461
ml.g5.48xlarge 2048 1 TRUE FALSE 418
ml.g5.48xlarge 1024 2 TRUE FALSE 423

Recommendations on instance types and hyperparameters

When fine-tuning the model’s accuracy, keep in mind the following:

  • Larger models such as 70B provide better performance than 8B
  • Performance without Int8 quantization is better than performance with Int8 quantization

Note the following training time and CUDA memory requirements:

  • Setting int8_quantization=True decreases the memory requirement and leads to faster training.
  • Decreasing per_device_train_batch_size and max_input_length reduces the memory requirement and therefore can be run on smaller instances. However, setting very low values may increase the training time.
  • If you’re not using Int8 quantization (int8_quantization=False), use FSDP (enable_fsdp=True) for faster and efficient training.

When choosing the instance type, consider the following:

  • At the time of writing this post, the G5 instances provided the most efficient training among the supported instance types. However, because AWS regularly updates and introduces new instance types, we recommend that you validate the recommended instance type for Meta Llama 3 fine-tuning in the SageMaker documentation or SageMaker console before proceeding.
  • Training time largely depends on the amount of GPUs and the CUDA memory available. Therefore, training on instances with the same number of GPUs (for example, ml.g5.2xlarge and ml.g5.4xlarge) is roughly the same. Therefore, you can use the more cost effective instance for training (ml.g5.2xlarge).

To learn about the cost of training per instance, refer to Amazon EC2 G5 Instances.

If your dataset is in instruction tuning format, where each sample consists of an instruction (input) and the desired model response (completion), and these input+completion sequences are short (for example, 50–100 words), using a high value for max_input_length can lead to poor performance. This is because the model may struggle to focus on the relevant information when dealing with a large number of padding tokens, and it can also lead to inefficient use of computational resources. The default value of -1 corresponds to a max_input_length of 1024 for Llama models. We recommend setting max_input_length to a smaller value (for example, 200–400) when working with datasets containing shorter input+completion sequences to mitigate these issues and potentially improve the model’s performance and efficiency.

Lastly, due to the high demand of the G5 instances, you may experience unavailability of these instances in your AWS Region with the error “CapacityError: Unable to provision requested ML compute capacity. Please retry using a different ML instance type.” If you experience this error, retry the training job or try a different Region.

Issues when fine-tuning large models

In this section, we discuss two issues when fine-tuning very large models.

Disable output compression

By default, the output of a training job is a trained model that is compressed in a .tar.gz format before it’s uploaded to Amazon S3. However, for large models like the 70B model, this compression step can be time-consuming, taking more than 4 hours. To mitigate this delay, it’s recommended to use the disable_output_compression feature supported by the SageMaker training environment. When disable_output_compression is set to True, the model is uploaded without any compression, which can significantly reduce the time taken for large model artifacts to be uploaded to Amazon S3. The uncompressed model can then be used directly for deployment or further processing. The following code shows how to pass this parameter into the SageMaker JumpStart estimator:

estimator = JumpStartEstimator(
model_id=model_id, environment={"accept_eula": "true"}, disable_output_compression=True
)

SageMaker Studio kernel timeout issue

Due to the size of the Meta Llama 3 70B model, the training job may take several hours to complete. The SageMaker Studio kernel is only used to initiate the training job, and its status doesn’t affect the ongoing training process. After the training job starts, the compute resources allocated for the job will continue running the training process, regardless of whether the SageMaker Studio kernel remains active or times out. If the kernel times out during the lengthy training process, you can still deploy the endpoint after training is complete using the training job name with the following code:

from sagemaker.jumpstart.estimator import JumpStartEstimator
training_job_name = <<<INSERT_TRAINING_JOB_NAME>>>

attached_estimator = JumpStartEstimator.attach(training_job_name, model_id)
attached_estimator.logs()
predictor = attached_estimator.deploy()

To find the training job name, navigate to the SageMaker console and under Training in the navigation pane, choose Training jobs. Identify the training job name and substitute it in the preceding code.

Clean up

To prevent incurring unnecessary charges, it’s recommended to clean up the deployed resources when you’re done using them. You can remove the deployed model with the following code:

predictor.delete_predictor()

Conclusion

In this post, we discussed fine-tuning Meta Llama 3 models using SageMaker JumpStart. We showed that you can use the SageMaker JumpStart console in SageMaker Studio or the SageMaker Python SDK to fine-tune and deploy these models. We also discussed the fine-tuning technique, instance types, and supported hyperparameters. In addition, we outlined recommendations for optimized training based on various tests we carried out.

The results for fine-tuning the three models over two datasets are shown in the appendix at the end of this post. As we can see from these results, fine-tuning improves summarization compared to non-fine-tuned models.

As a next step, you can try fine-tuning these models on your own dataset using the code provided in the GitHub repository to test and benchmark the results for your use cases.


About the Authors

Ben FriebeBen Friebe is a Senior Solutions Architect at Amazon Web Services, based in Brisbane, Australia. He likes computers.

Pavan Kumar Rao Navule<Pavan Kumar Rao Navule is a Solutions Architect at Amazon Web Services, where he works with ISVs in India to help them innovate on the AWS platform. He is specialized in architecting AI/ML and generative AI services at AWS. Pavan is a published author for the book “Getting Started with V Programming.” In his free time, Pavan enjoys listening to the great magical voices of Sia and Rihanna.

Khush PatelKhush Patel Khush Patel is a Solutions Architect at Amazon Web Services based out of Houston, Texas. He’s passionate about working with customers to deliver business value using technology. He has a multitude of experience with customers working with Machine Learning and GenerativeAI workloads. In his free time, Khush enjoys watching sports and reading.

Dr. Farooq Sabir is a Senior Artificial Intelligence and Machine Learning Specialist Solutions Architect at AWS. He holds PhD and MS degrees in Electrical Engineering from the University of Texas at Austin and an MS in Computer Science from Georgia Institute of Technology. He has over 15 years of work experience and also likes to teach and mentor college students. At AWS, he helps customers formulate and solve their business problems in data science, machine learning, computer vision, artificial intelligence, numerical optimization, and related domains. Based in Dallas, Texas, he and his family love to travel and go on long road trips.


Appendix

This appendix provides additional information about performance benchmarking and dataset formatting.

Performance benchmarking

In this section, we provide results for fine-tuning the two Meta Llama 3 models (8B and 70B) on two different datasets: Dolly and Dialogsum. For the Dolly dataset, our task is to summarize a paragraph of text, whereas for Dialogsum, we are fine-tuning the model to summarize a discussion between two people. In the following tables, we show the input to the model (prompt and instructions), ground truth (summary), response from the pre-trained Meta Llama 3 model, and response from the fine-tuned Meta Llama 3 model for each of the models. We show inference results for five data points. You can notice from the following tables that the summaries improve for both the datasets when we fine-tune the models.

Results for fine-tuning the Meta Llama 3 8B text generation model on the Dolly dataset

Inputs Ground Truth Response from Non-Fine-Tuned Model Response from Fine-Tuned Model
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nCan you generate a 2 lines abstract of this text?\n\n### Input:\nIn physics, gravity (from Latin gravitas ‘weight’) is a fundamental interaction which causes mutual attraction between all things with mass or energy. Gravity is, by far, the weakest of the four fundamental interactions, approximately 1038 times weaker than the strong interaction, 1036 times weaker than the electromagnetic force and 1029 times weaker than the weak interaction. As a result, it has no significant influence at the level of subatomic particles. However, gravity is the most significant interaction between objects at the macroscopic scale, and it determines the motion of planets, stars, galaxies, and even light.\n\nOn Earth, gravity gives weight to physical objects, and the Moon’s gravity is responsible for sublunar tides in the oceans (the corresponding antipodal tide is caused by the inertia of the Earth and Moon orbiting one another). Gravity also has many important biological functions, helping to guide the growth of plants through the process of gravitropism and influencing the circulation of fluids in multicellular organisms. Investigation into the effects of weightlessness has shown that gravity may play a role in immune system function and cell differentiation within the human body.\n\nThe gravitational attraction between the original gaseous matter in the universe allowed it to coalesce and form stars which eventually condensed into galaxies, so gravity is responsible for many of the large-scale structures in the universe. Gravity has an infinite range, although its effects become weaker as objects get farther away.\n\nGravity is most accurately described by the general theory of relativity (proposed by Albert Einstein in 1915), which describes gravity not as a force, but as the curvature of spacetime, caused by the uneven distribution of mass, and causing masses to move along geodesic lines. The most extreme example of this curvature of spacetime is a black hole, from which nothing—not even light—can escape once past the black hole’s event horizon However, for most applications, gravity is well approximated by Newton’s law of universal gravitation, which describes gravity as a force causing any two bodies to be attracted toward each other, with magnitude proportional to the product of their masses and inversely proportional to the square of the distance between them.\n\n\n\n### Response:\n In general physics theory (Newton), Gravity is a mutual attraction between objects based on their respective masses and the square of the distance which separates them. In the relativity theory (Einstein), it’s a curvature of the spacetime generated by the objects themselves. Gravity is a fundamental interaction which causes mutual attraction between all things with mass or energy. It is the weakest of the four fundamental interactions, but it is the most significant interaction between objects at the macroscopic scale, and it determines the motion of planets, stars, galaxies, and even light. Gravity gives weight to physical objects, and the Moon’s gravity is responsible for sublunar tides in the oceans. It also has many important biological functions, helping to guide the growth of plants through the process Gravity is a fundamental interaction which causes mutual attraction between all things with mass or energy. Gravity is, by far, the weakest of the four fundamental interactions, approximately 1038 times weaker than the strong interaction, 1036 times weaker than the electromagnetic force and 1029 times weaker than the weak interaction. Gravity is most accurately described by the general theory of relativity (proposed by Albert Einstein in 1915), which describes gravity not as a force, but as the curvature of spacetime
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nWhat is ProSavin\n\n### Input:\nProSavin is an experimental drug believed to be of use in the treatment of Parkinson’s disease. It is administered to the striatum in the brain, inducing production of dopamine.\n\nIt is manufactured by Oxford BioMedica. Results from a Phase I/II clinical trial were published in the Lancet and showed safety, but little efficacy. ProSavin was superseded by AXO-Lenti-PD (OXB-102), an optimized version of the drug.\n\n\n\n### Response:\n ProSavin is an experimental drug believed to be of use in the treatment of Parkinson’s disease. It is administered to the striatum in the brain, inducing production of dopamine. ProSavin is an experimental drug believed to be of use in the treatment of Parkinson’s disease. It is administered to the striatum in the brain, inducing production of dopamine.\n\nIt is manufactured by Oxford BioMedica. Results from a Phase I/II clinical trial were published in the Lancet and showed safety, but little efficacy. ProSavin was superseded by AXO-Lenti-PD (OXB-102), an optimized version of the drug. ProSavin is an experimental drug believed to be of use in the treatment of Parkinson’s disease. It is administered to the striatum in the brain, inducing production of dopamine.\n\nIt is manufactured by Oxford BioMedica. Results from a Phase I/II clinical trial were published in the Lancet and showed safety, but little efficacy. ProSavin was superseded by AXO-Lenti-PD (OXB-102), an optimized version of the drug.
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nWho were the Beatles?\n\n### Input:\nThe Beatles were an English rock band, formed in Liverpool in 1960, that comprised John Lennon, Paul McCartney, George Harrison and Ringo Starr. They are regarded as the most influential band of all time and were integral to the development of 1960s counterculture and popular music’s recognition as an art form. Rooted in skiffle, beat and 1950s rock ‘n’ roll, their sound incorporated elements of classical music and traditional pop in innovative ways; the band also explored music styles ranging from folk and Indian music to psychedelia and hard rock. As pioneers in recording, songwriting and artistic presentation, the Beatles revolutionised many aspects of the music industry and were often publicised as leaders of the era’s youth and sociocultural movements.\n\n\n\n### Response:\n The Beatles were a rock band comprised of 4 musicians. They were pioneers who had a huge influence on the music industry and wider sociocultural movements. The Beatles were an English rock band, formed in Liverpool in 1960, that comprised John Lennon, Paul McCartney, George Harrison and Ringo Starr. They are regarded as the most influential band of all time and were integral to the development of 1960s counterculture and popular music’s recognition as an art form. Rooted in skiffle, beat and 1950s rock ‘n’ roll, their sound incorporated elements of classical music and traditional pop in innovative ways; the The Beatles were an English rock band, formed in Liverpool in 1960, that comprised John Lennon, Paul McCartney, George Harrison and Ringo Starr. They are regarded as the most influential band of all time and were integral to the development of 1960s counterculture and popular music’s recognition as an art form. Rooted in skiffle, beat and 1950s rock ‘n’ roll, their sound incorporated elements of classical music and traditional pop in innovative ways; the
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nSummarize what Alpha Phi is.\n\n### Input:\nAlpha Phi International Women’s Fraternity (ΑΦ, also known as APhi) is an international sorority with 172 active chapters and over 250,000 initiated members.\n\nFounded at Syracuse University in Syracuse, New York on September 18, 1872, it is the fourth Greek-letter organization founded for women, and the first women’s fraternity founded in the northeast.\n\nAlpha Phi is a member of the National Panhellenic Conference, the governing council of 26 women’s fraternities. Alpha Phi’s international headquarters are located in Evanston, Illinois.\n\n\n\n### Response:\n Alpha Phi is one of the first women’s fraternities founded in the Northeast. They have over 172 active chapters at College campuses and over 250,000 initiated members. Alpha Phi was founded at Syracuse University in Syracuse, New York on September 18, 1872. Alpha Phi is an international sorority with 172 active chapters and over 250,000 initiated members. Founded at Syracuse University in Syracuse, New York on September 18, 1872, it is the fourth Greek-letter organization founded for women, and the first women’s fraternity founded in the northeast. Alpha Phi is a member of the National Panhellenic Conference, the governing council of 26 women’s fraternities. Alpha Phi’s international headquarters are located in Evanston, Illinois Alpha Phi is an international sorority with 172 active chapters and over 250,000 initiated members. It was founded at Syracuse University in Syracuse, New York on September 18, 1872, and is the fourth Greek-letter organization founded for women, and the first women’s fraternity founded in the northeast. Alpha Phi is a member of the National Panhellenic Conference, the governing council of 26 women’s fraternities. Alpha Phi’s international headquarters are located in Evanston
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nplease give me a short description about SVB collapse\n\n### Input:\nOn Friday, March 10, 2023, Silicon Valley Bank (SVB) failed after a bank run, marking the second-largest bank failure in United States history and the largest since the 2007–2008 financial crisis. It was one of three March 2023 United States bank failures.\n\n\n\n### Response:\n It’s due to risk mismanagement of SVB Bank, where the bank bet on long dated treasury bond. When the bank runs happen, SVB don’t have a choice but to sell those long dated bond at a steep discount and incur huge lossses. Silicon Valley Bank (SVB) was a bank that failed in 2023. It was one of the largest banks in the United States, and its failure had a significant impact on the financial markets. The bank was founded in 1983 and was headquartered in Santa Clara, California. It was known for its focus on the technology industry and its support of startups and venture capital firms. However, in 2023, the bank faced a number of challenges, including a decline in the value Silicon Valley Bank (SVB) failed after a bank run, marking the second-largest bank failure in United States history and the largest since the 2007–2008 financial crisis. It was one of three March 2023 United States bank failures.

Results for fine-tuning the Meta Llama 3 8B text generation model on the Dialogsum dataset

Inputs Ground Truth Response from Non-Fine-Tuned Model Response from Fine-Tuned Model
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: Good moring, madam. Can I help you? \n#Person2#: Good moring. I am Sunfang. Yesterday Mr. Smith told me to come for a job interview this morning. Is he available now? \n#Person1#: Yes, he is waiting for you in Room 123. \n#Person2#: Thank you very much. May I come in? \n#Person3#: Come in, please. \n#Person2#: Good moring, Mr. Smith. I am Sunfang coming for an interview. \n#Person3#: Nice to see you. Take a seat, please. \n#Person2#: Thank you. \n#Person3#: We’ve been looking over your application. I see you’ve got several years experience in secretarial work. What kind of work did you do in your previous job? \n#Person2#: I mainly did typing, filing, answering telephone calls and such routine work in the beginning. I started the job as a junior secretary, but within the year I was promoted to senior secretary of Southeast Company. I drafted business correspondence for the general manager. \n#Person3#: Then what’s the reason that promoted you to leave Southeast Company? \n#Person2#: Last month the company moved to Guangzhou for further development. I was reluctant to relocate because I love Nanjing so much. \n#Person3#: Tell me what you know about our company. \n#Person2#: Your company is the leading dealer in the import and export of cosmetics. Also, you have a reputation for being a good company to work for and you are in last year’s top 500 Employers list. \n#Person3#: Tell me three positive qualities your coworkers would use to describe you and your work style. \n#Person2#: First they’d say I am very punctual. They’d also say I am a cheerful person to work with I don’t bring my problems to work. I try to have a positive attitude about work. And thirdly, they’d say I am flexible I accept change easily. \n#Person3#: Tell me something about your education. \n#Person2#: I graduated from Yangzhou Junior College. Now I am attending evening classes in Nanjing Radio & TV University three times a week. My major is International Trade. I’ve already passed over half of the compulsory courses and five optional courses, and expect to receive a B. A. degree next year. \n#Person3#: I see you are an ambitious young lady. I wish you can make greater progress in the future. \n\n\n\n### Response:\n #Person1# leads Sunfang to Mr. Smith for an interview. Sunfang tells Mr. Smith she was promoted fast in her previous job. She wants to work for Mr. Smith because her old company moved but she didn’t want to relocate, and because She thinks Mr. Smith’s company plays a leading role. Sunfang thinks her colleagues might describe her as being punctual, positive and flexible. Mr. Smith is impressed. #Person1#: Good moring, madam. Can I help you? \n#Person2#: Good moring. I am Sunfang. Yesterday Mr. Smith told me to come for a job interview this morning. Is he available now? \n#Person1#: Yes, he is waiting for you in Room 123. \n#Person2#: Thank you very much. May I come in? \n#Person3#: Come in, please. \n#Person2#: Good moring, Mr. Smith. I am Sunfang coming for an interview. \n#Person3#: Nice to see you. Take a seat, please. \n#Person2#: Thank you. \n#Person3#: We’ve been looking over your application. I see you’ve got several years experience in secretarial work. What kind of work did you do in your previous job? \n#Person2#: I mainly did typing, filing, answering telephone calls and such routine work in the beginning. I started the job as a junior secretary, but within the year I was promoted to senior secretary of Southeast Company. I drafted business correspondence for the general manager. \n#Person3#: Then what’s the reason that promoted you to leave Southeast Company? \n#Person2#: Last month the company moved to Guangzhou for further development. I was reluctant to relocate because I love Nanjing so much. \n#Person3#: Tell me what you know about our company. \n#Person2#: Your company Sunfang comes for a job interview with Mr. Smith. She tells Mr. Smith about her previous job and her work experience. She also tells Mr. Smith about her company and her education.
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: Well, Jenny, the school year is almost over. We just have two more weeks before exams. What do you plan to do this summer?\n#Person2#: I’m going to teach English to some immigrants in the university’s community service program.\n#Person1#: That sounds interesting. Don’t you need to speak a foreign language for a job like that?\n#Person2#: No, you don’t. You just have to present the language simply and give the students a chance to practice speaking.\n#Person1#: Come to think of it, that’s the way I was taught to speak Chinese. But speaking didn’t help me learn to read and write Chinese.\n#Person2#: My students won’t want to read and write English, at lease not now. They are more interested in speaking.\n#Person1#: You sound very knowledgeable about all this. How do you know so much?\n#Person2#: I took a Teaching English as a Second Language course last year when you were in China. I’ve also talked with the experienced teachers quite a lot. I think I would like to be an ESL teacher when I graduate.\n\n\n\n### Response:\n Jenny plans to teach English to some immigrants in the university’s community service program this summer and she tells #Person1# that students are more interested in speaking. #Person1#: Well, Jenny, the school year is almost over. We just have two more weeks before exams. What do you plan to do this summer?\n#Person2#: I’m going to teach English to some immigrants in the university’s community service program.\n#Person1#: That sounds interesting. Don’t you need to speak a foreign language for a job like that?\n#Person2#: No, you don’t. You just have to present the language simply and give the students a chance to practice speaking.\n#Person1#: Come to think of it, that’s the way I was taught to speak Chinese. But speaking didn’t help me learn to read and write Chinese.\n#Person2#: My students won’t want to read and write English, at lease not now. They are more interested in speaking.\n#Person1#: You sound very knowledgeable about all this. How do you know so much?\n#Person2#: I took a Teaching English as a Second Language course last year when you were in China. I’ve also talked with the experienced teachers quite a lot. I think I would like to be an ESL teacher when I graduate.\n Jenny tells #Person1# she’s going to teach English to some immigrants in the university’s community service program. Jenny thinks speaking English is the most important part and she’s knowledgeable about it.
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: Can I help you?\n#Person2#: Yes, I am looking for a pair of gloves. I think I wear size six.\n#Person1#: The gloves are on this counter. Let’s see… How about the blue pair? The color suits you and they are washable, too.\n#Person2#: Oh, I’ll try them on. They seem to fit. How much shall I pay you?\n#Person1#: They are on sales this week. They have been reduced to twenty-five dollars.\n#Person2#: All right. I will take them.\n#Person1#: They are very nice.\n#Person2#: Thank you.\n\n\n\n### Response:\n #Person2# buys a pair of blue gloves of size six with #Person1#’s assistance. #Person1#: Can I help you?\n#Person2#: Yes, I am looking for a pair of gloves. I think I wear size six.\n#Person1#: The gloves are on this counter. Let’s see… How about the blue pair? The color suits you and they are washable, too.\n#Person2#: Oh, I’ll try them on. They seem to fit. How much shall I pay you?\n#Person1#: They are on sales this week. They have been reduced to twenty-five dollars.\n#Person2#: All right. I will take them.\n#Person1#: They are very nice.\n#Person2#: Thank you.\n #Person2# buys a pair of gloves with #Person1#’s help.
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: How is your business on the Internet, Steven?\n#Person2#: Don’t mention it. I lost a bunch of money on it this year. I think it is on the brink of bankruptcy.\n#Person1#: I am sorry to hear that. It shows that you should buy companies, not ideas.\n#Person2#: I changed my strategy around too. I am investing in good companies now.\n#Person1#: Yes, I agree with you. I am going to find a good company and invest in it soon.\n#Person2#: It’s a good idea. But please be careful when you invest this time.\n#Person1#: Yes, I will. I should learn more from you. After all I am green hand on it.\n#Person2#: If you have any question, you can ask me.\n#Person1#: OK. My wife sent me a message and asked me to go home just now. I am leaving.\n#Person2#: All right. Let’s chat next time. Bye-bye.\n#Person1#: Bye.\n\n\n\n### Response:\n Steven lost much money on Internet business so he’s now investing in good companies. #Person1# wants to invest too but #Person1# is green hand on investment. Steven’s willing to help. #Person1#: How is your business on the Internet, Steven?\n#Person2#: Don’t mention it. I lost a bunch of money on it this year. I think it is on the brink of bankruptcy.\n#Person1#: I am sorry to hear that. It shows that you should buy companies, not ideas.\n#Person2#: I changed my strategy around too. I am investing in good companies now.\n#Person1#: Yes, I agree with you. I am going to find a good company and invest in it soon.\n#Person2#: It’s a good idea. But please be careful when you invest this time.\n#Person1#: Yes, I will. I should learn more from you. After all I am green hand on it.\n#Person2#: If you have any question, you can ask me.\n#Person1#: OK. My wife sent me a message and asked me to go home just now. I am leaving.\n#Person2#: All right. Let’s chat next time. Bye-bye.\n#Person1#: Bye.\n Steven lost a bunch of money on the Internet and changed his strategy to invest in good companies. #Person1# agrees and will be careful when investing.
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: Which language do you speak?\n#Person2#: Well, I studied French at school, but I’ve forgotten most of it. I learnt some Russian while I was studying there and now I study Spanish in evening classes. Do you speak Japanese at home with your parents.\n#Person1#: Yes, I do. I’m learning Chinese now on the internet.\n#Person2#: How do you find studying a language on the internet?\n#Person1#: It’s not easy, particularly with listening and speaking, because sometimes the connection isn’t too good.\n#Person2#: I bet you find the character difficult to write!\n#Person1#: Actually, many of the characters are the same or very similar in both Chinese and Japanese, so they’re easy for me.\n#Person2#: So, you can read and write much better than you speak and understand other speakers, right?\n#Person1#: Yes. How are the Spanish classes going?\n#Person2#: In a way, it’s similar to your experience of learning Chinese. Some Spanish words are very similar to English ones. But I need to practise listening to native speakers and saying the words myself.\n\n\n\n### Response:\n #Person1# is learning Chinese on the internet and finds listening and speaking difficult, but thinks writing is easy. #Person2# thinks #Person2#’s experience of learning Spanish is similar to #Person1#’s. #Person1#: Which language do you speak?\n#Person2#: Well, I studied French at school, but I’ve forgotten most of it. I learnt some Russian while I was studying there and now I study Spanish in evening classes. Do you speak Japanese at home with your parents.\n#Person1#: Yes, I do. I’m learning Chinese now on the internet.\n#Person2#: How do you find studying a language on the internet?\n#Person1#: It’s not easy, particularly with listening and speaking, because sometimes the connection isn’t too good.\n#Person2#: I bet you find the character difficult to write!\n#Person1#: Actually, many of the characters are the same or very similar in both Chinese and Japanese, so they’re easy for me.\n#Person2#: So, you can read and write much better than you speak and understand other speakers, right?\n#Person1#: Yes. How are the Spanish classes going?\n#Person2#: In a way, it’s similar to your experience of learning Chinese. Some Spanish words are very similar to English ones. But I need to practise listening to native speakers and saying the words myself.\n #Person1# and #Person2# are talking about the languages they speak. #Person1# is learning Chinese on the internet and #Person2# is learning Spanish. They both find it difficult to learn the language.

Results for fine-tuning the Meta Llama 3 70B text generation model on the Dolly dataset

Inputs Ground Truth Response from Non-Fine-Tuned Model Response from Fine-Tuned Model
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nCan you give me a short summary of the follow paragraph about the history of electronic music?\n\n### Input:\nDuring the 1960s, digital computer music was pioneered, innovation in live electronics took place, and Japanese electronic musical instruments began to influence the music industry. In the early 1970s, Moog synthesizers and drum machines helped popularize synthesized electronic music. The 1970s also saw electronic music begin to have a significant influence on popular music, with the adoption of polyphonic synthesizers, electronic drums, drum machines, and turntables, through the emergence of genres such as disco, krautrock, new wave, synth-pop, hip hop, and EDM. In the early 1980s mass-produced digital synthesizers, such as the Yamaha DX7, became popular, and MIDI (Musical Instrument Digital Interface) was developed. In the same decade, with a greater reliance on synthesizers and the adoption of programmable drum machines, electronic popular music came to the fore. During the 1990s, with the proliferation of increasingly affordable music technology, electronic music production became an established part of popular culture. In Berlin starting in 1989, the Love Parade became the largest street party with over 1 million visitors, inspiring other such popular celebrations of electronic music.\n\n\n\n### Response:\n With the advent of new electronic instruments in the 60s and 70s, namely synthesizers and drum machines, electronic music quickly gained mainstream popularity through genres like disco and hip hop. As the technology progressed through the 80s and 90s, more affordable instruments led electronic music to be even more widely embraced, cementing its status in pop culture and leading to the rise of large electronic music events around the globe. The history of electronic music can be traced back to the 1960s, when digital computer music was pioneered, innovation in live electronics took place, and Japanese electronic musical instruments began to influence the music industry. In the early 1970s, Moog synthesizers and drum machines helped popularize synthesized electronic music. The 1970s also saw electronic music begin to have a significant influence on popular music, with the adoption of polyphonic synthesizers, electronic drums, drum machines, and Electronic music has been around since the 1960s. In the 1970s, Moog synthesizers and drum machines helped popularize synthesized electronic music. In the 1980s, digital synthesizers became popular and MIDI was developed. In the 1990s, electronic music production became an established part of popular culture.
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nWho was Nike in Greek Mythology?\n\n### Input:\nIn Greek mythology, Nike (/ˈnaɪki/ (listen); Ancient Greek: Νίκη, lit. ‘victory’, ancient: [nǐː.kɛː], modern: [ˈni.ci]) was a goddess who personified victory in any field including art, music, war, and athletics. She is often portrayed in Greek art as Winged Victory in the motion of flight; however, she can also appear without wings as “Wingless Victory” when she is being portrayed as an attribute of another deity such as Athena. In Greek literature Nike is described as both an attribute and attendant to the gods Zeus and Athena. Nike gained this honored role beside Zeus during the Titanomachy where she was one of the first gods to offer her allegiance to Zeus. At Athens, Nike became a servant to Athena as well as an attribute of her due to the prominent status Athena held in her patron city. The fusion of the two goddesses at Athens has contributed to the ambiguity surrounding Nike’s origins. It is unclear whether she originated from a character trait of the Greek goddess Athena or has always existed as an independent deity. Her origin story in Greek mythology is also slightly ambiguous, with the Theogony claiming Nike to be the daughter of Styx and Pallas while the Homeric Hymns describe Ares as being Nike’s father. Her Roman equivalent was Victoria.\n\n\n\n### Response:\n Nike is the god of ‘victory’ in Greek mythology, and is often portrayed in art Winged Victory in the motion of flight. Nike was the Greek goddess of victory. She was the daughter of the Titan Pallas and the goddess Styx, and the sister of Kratos (Strength), Bia (Force), and Zelus (Rivalry). She was depicted as a winged goddess, and was often shown holding a wreath of victory or a palm branch. She was also associated with the goddess Athena, and was often depicted as her companion. Nike was worshipped throughout Greece, and her cult was particularly strong Nike was the Greek goddess of victory. She was often portrayed in Greek art as Winged Victory in the motion of flight. Nike gained this honored role beside Zeus during the Titanomachy where she was one of the first gods to offer her allegiance to Zeus. At Athens, Nike became a servant to Athena as well as an attribute of her due to the prominent status Athena held in her patron city.
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nGive me some details about the MV Cape Race please\n\n### Input:\nThe MV Cape Race (T-AKR-9960) was originally built as the MV Seaspeed America for the Transocean RO-RO Corp. She was sold to various other corporations until she was purchased by the United States Navy in April 1993. She underwent conversion to fit a more active role in and serve a military need better and was placed back in service April 1993. She has served on multiple occasions but is currently in the US Ready Reserve Fleet.\n\n\n\n### Response:\n The MV Cape Race (T-AKR-9960) was originally built as the MV Seaspeed America for the Transocean RO-RO Corp. She was sold to various other corporations until she was purchased by the United States Navy in April 1993. \n\nShe underwent conversion to fit a more active role in and serve a military need better and was placed back in service April 1993. She has served on multiple occasions but is currently in the US Ready Reserve Fleet. The MV Cape Race (T-AKR-9960) was originally built as the MV Seaspeed America for the Transocean RO-RO Corp. She was sold to various other corporations until she was purchased by the United States Navy in April 1993. She underwent conversion to fit a more active role in and serve a military need better and was placed back in service April 1993. She has served on multiple occasions but is currently in the US Ready Reserve Fleet. The MV Cape Race (T-AKR-9960) was originally built as the MV Seaspeed America for the Transocean RO-RO Corp. She was sold to various other corporations until she was purchased by the United States Navy in April 1993. She underwent conversion to fit a more active role in and serve a military need better and was placed back in service April 1993. She has served on multiple occasions but is currently in the US Ready Reserve Fleet.
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nWhat is sweep rowing?\n\n### Input:\nRowing, sometimes called crew in the United States, is the sport of racing boats using oars. It differs from paddling sports in that rowing oars are attached to the boat using oarlocks, while paddles are not connected to the boat. Rowing is divided into two disciplines: sculling and sweep rowing. In sculling, each rower holds two oars—one in each hand, while in sweep rowing each rower holds one oar with both hands. There are several boat classes in which athletes may compete, ranging from single sculls, occupied by one person, to shells with eight rowers and a coxswain, called eights. There are a wide variety of course types and formats of racing, but most elite and championship level racing is conducted on calm water courses 2 kilometres (1.2 mi) long with several lanes marked using buoys.\n\n\n\n### Response:\n Sweep rowing is a rowing discipline in which rowers hold one oar with both hands. Sweep rowing is a type of rowing where each rower holds one oar with both hands. There are several boat classes in which athletes may compete, ranging from single sculls, occupied by one person, to shells with eight rowers and a coxswain, called eights. There are a wide variety of course types and formats of racing, but most elite and championship level racing is conducted on calm water courses 2 kilometres (1.2 mi) long with several lanes Sweep rowing is a type of rowing where each rower holds one oar with both hands.
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nWhat is the Pandemis corylana\n\n### Input:\nPandemis corylana, the chequered fruit-tree tortrix, hazel tortrix moth, filbert tortricid or barred fruit tree moth, is a moth of the family Tortricidae. It is found from northern and central Europe to Siberia, Korea and Japan.\n\n\n\n### Response:\n The Pandemis corylana, the chequered fruit-tree tortrix, hazel tortrix moth, filbert tortricid or barred fruit tree moth, is a moth of the family Tortricidae. It is found from northern and central Europe to Siberia, Korea and Japan. Pandemis corylana, the chequered fruit-tree tortrix, hazel tortrix moth, filbert tortricid or barred fruit tree moth, is a moth of the family Tortricidae. It is found from northern and central Europe to Siberia, Korea and Japan. The Pandemis corylana, also known as the chequered fruit-tree tortrix, hazel tortrix moth, filbert tortricid or barred fruit tree moth, is a moth of the family Tortricidae. It is found from northern and central Europe to Siberia, Korea and Japan.

Results for fine-tuning the Meta Llama 3 70B text generation model on the Dialogsum dataset

Inputs Ground Truth Response from Non-Fine-Tuned Model Response from Fine-Tuned Model
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: Would you like some more coffee, James?\n#Person2#: Oh, yes, please. Another day’s work. I like the job, but I hate traveling around London.\n#Person1#: Oh, come on, James. How do you get to work?\n#Person2#: I walk to the station and take the underground. And sometimes I take the taxi if I get up late.\n#Person1#: How long does the journey take?\n#Person2#: It takes about 45 minutes on a good day. It’s just awful.\n#Person1#: Well, how much does it cost?\n#Person2#: It’s twenty pounds a week. I think that’s quite a lot.\n#Person1#: Oh, James. Stop complaining. And hurry up. You’ll be late.\n\n\n\n### Response:\n James likes the job but hates the commute because it takes a long time and costs a lot. #Person1#: Would you like some more coffee, James?\n#Person2#: Oh, yes, please. Another day’s work. I like the job, but I hate traveling around London.\n#Person1#: Oh, come on, James. How do you get to work?\n#Person2#: I walk to the station and take the underground. And sometimes I take the taxi if I get up late.\n#Person1#: How long does the journey take?\n#Person2#: It takes about 45 minutes on a good day. It’s just awful.\n#Person1#: Well, how much does it cost?\n#Person2#: It’s twenty pounds a week. I think that’s quite a lot.\n#Person1#: Oh, James. Stop complaining. And hurry up. You’ll be late.\n\n#Person1# asked #Person2# if he would like some more coffee. #Person2# said yes and that he liked his job but hated traveling around London. #Person1# asked how he got to work. #Person2# said he walked to the station and took the underground, and sometimes took a taxi if he got up late. #Person1# asked how long the journey took. #Person2# said it took about 45 minutes on a good day and that it was awful. #Person1# asked how much it cost. #Person2# said it was twenty pounds a week and that he thought that was quite a lot James complains to #Person1# about the traveling around London. #Person1# asks him to hurry up.
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: What is so special about tomorrow?\n#Person2#: Tomorrow is October 24th, United Nations Day.\n#Person1#: Why is October 24th United Nations Day?\n#Person2#: That’s the date in 1945 when the United Nations began.\n#Person1#: How does your school celebrate United Nations Day?\n#Person2#: The UN Day celebration includes the songs and dances of different countries in the United Nations and a traditional dress parade. Our school library displays students’ works of art from different countries and cultures. We also hold a ‘Tastes from Around the World’ food festival.\n#Person1#: What’s the purpose of the celebration?\n#Person2#: To help us students learn about other cultures and understand and respect other cultures.\n#Person1#: It is really worthwhile.\n\n\n\n### Response:\n #Person1# and #Person2# talk about the United Nations Day. #Person2# tells #Person1# how the school will celebrate the day and the purpose of it. #Person1#: What is so special about tomorrow?\n#Person2#: Tomorrow is October 24th, United Nations Day.\n#Person1#: Why is October 24th United Nations Day?\n#Person2#: That’s the date in 1945 when the United Nations began.\n#Person1#: How does your school celebrate United Nations Day?\n#Person2#: The UN Day celebration includes the songs and dances of different countries in the United Nations and a traditional dress parade. Our school library displays students’ works of art from different countries and cultures. We also hold a ‘Tastes from Around the World’ food festival.\n#Person1#: What’s the purpose of the celebration?\n#Person2#: To help us students learn about other cultures and understand and respect other cultures.\n#Person1#: It is really worthwhile.\n #Person2# tells #Person1# United Nations Day is October 24th. #Person2#’s school celebrates it to help students understand and respect other cultures.
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: Have you heard about Alice and David?\n#Person2#: No, have they had another fight?\n#Person1#: No, they’ve got engaged.\n#Person2#: You must be joking. Those two?\n#Person1#: Well, my dear, I didn’t believe it either. But I got it straight from the horse’s mouth. David called me this morning.\n#Person2#: So when did this all happen?\n#Person1#: Last weekend, while they were on their ski trip.\n#Person2#: Well, I believe it now, and when are they are getting married?\n#Person1#: Next June.\n#Person2#: I can hardly believe it.\n\n\n\n### Response:\n #Person1# and #Person2# can’t believe that Alice and David, who fought with each other, got engaged. #Person1# told #Person2# that Alice and David had got engaged. #Person2# was surprised and asked when it had happened. #Person1# said that it had happened last weekend, while they were on their ski trip. #Person2# asked when they were getting married. #Person1# said that they were getting married next June. #Person1# tells #Person2# that Alice and David have got engaged. #Person2# is surprised.
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: We did a survey on your job performance.\n#Person2#: How did I do?\n#Person1#: Not so good. Lots of complains.\n#Person2#: Really? I thought I was doing well.\n#Person1#: Your performance so far is not quite in line with what we expect for our executives. I hope I’Ve made my point.\n\n\n\n### Response:\n #Person1# tells #Person2# #Person2#’s job performance isn’t satisfying, despite #Person2#’s opposite opinion. #Person1# told #Person2# that they did a survey on their job performance. #Person2# asked how they did, and #Person1# said that they did not do so good and that there were lots of complains. #Person2# was surprised and said that they thought they were doing well. #Person1# said that their performance so far was not quite in line with what they expect for their executives, and they hoped that they had made their point. #Person1# tells #Person2# that #Person2#’s job performance is not good.
Below is a Instruction that holds conversation which describes discussion between two people.Write a response that appropriately summarizes the conversation.\n\n### Instruction:\n#Person1#: Hi, Francis, morning! How are you doing today?\n#Person2#: I had a bad start of the day. I was stuck in a bad traffic jam on my way and I missed my morning meeting.\n#Person1#: Indeed, I agree that is not a good start. How about the meeting? You missed anything important?\n#Person2#: It is ok. My boss briefs me later on.\n#Person1#: Was he unhappy with you not attending the meeting?\n#Person2#: Err, kind of. You know this traffic jam is really bad with so many new cars hitting the streets these days.\n\n\n\n### Response:\n Francis was stuck in a traffic jam and missed the morning meeting. His boss briefs him. #Person1#: Hi, Francis, morning! How are you doing today?\n#Person2#: I had a bad start of the day. I was stuck in a bad traffic jam on my way and I missed my morning meeting.\n#Person1#: Indeed, I agree that is not a good start. How about the meeting? You missed anything important?\n#Person2#: It is ok. My boss briefs me later on.\n#Person1#: Was he unhappy with you not attending the meeting?\n#Person2#: Err, kind of. You know this traffic jam is really bad with so many new cars hitting the streets these days.\n\n#Person1# asked #Person2# how he was doing today. #Person2# said he had a bad start of the day because he was stuck in a bad traffic jam on his way and he missed his morning meeting. #Person1# agreed that it was not a good start and asked if #Person2# missed anything important in the meeting. #Person2# said it was ok because his boss would brief him later on. #Person1# asked if his boss was unhappy with him not attending the meeting. #Person2# said he was kind of unhappy because the traffic jam was really bad with so many new cars hitting the streets these days. Francis tells #Person1# he missed the meeting because of the traffic jam.

Dataset formatting

We currently offer two types of fine-tuning: instruction fine-tuning and domain adaption fine-tuning. You can switch to one of the training methods by specifying the parameter instruction_tuned as True or False.

Domain adaption format

The text generation model can be fine-tuned on any domain-specific dataset to incorporate domain-specific knowledge and language patterns. After fine-tuning on the domain-specific dataset, the model is expected to generate more relevant and accurate text within that domain. Although few-shot prompting can also guide the model towards domain-specific generation, the fine-tuning process plays a crucial role in adapting the model’s understanding and generation capabilities to the target domain. The combination of fine-tuning on domain data and effective prompting techniques can enable the model to perform various NLP tasks within that specific domain more effectively.

For input to the model, use a training and optional validation directory. Each directory contains a CSV, JSON, or TXT file. For CSV and JSON files, the train or validation data is used from the column called text or the first column if no column called text is found. The number of files under train and validation (if provided) should equal to 1, respectively.

The output is a trained model that can be deployed for inference.

The following is an example of a TXT file for fine-tuning the text generation model. The TXT file is SEC filings of Amazon from 2021–2022:

This report includes estimates, projections, statements relating to our
business plans, objectives, and expected operating results that are “forward-
looking statements” within the meaning of the Private Securities Litigation
Reform Act of 1995, Section 27A of the Securities Act of 1933, and Section 21E
of the Securities Exchange Act of 1934. Forward-looking statements may appear
throughout this report, including the following sections: “Business” (Part I,
Item 1 of this Form 10-K), “Risk Factors” (Part I, Item 1A of this Form 10-K),
and “Management’s Discussion and Analysis of Financial Condition and Results
of Operations” (Part II, Item 7 of this Form 10-K). These forward-looking
statements generally are identified by the words “believe,” “project,”
“expect,” “anticipate,” “estimate,” “intend,” “strategy,” “future,”
“opportunity,” “plan,” “may,” “should,” “will,” “would,” “will be,” “will
continue,” “will likely result,” and similar expressions. Forward-looking
statements are based on current expectations and assumptions that are subject
to risks and uncertainties that may cause actual results to differ materially.
We describe risks and uncertainties that could cause actual results and events
to differ materially in “Risk Factors,” “Management’s Discussion and Analysis
of Financial Condition and Results of Operations,” and “Quantitative and
Qualitative Disclosures about Market Risk” (Part II, Item 7A of this Form
10-K). Readers are cautioned not to place undue reliance on forward-looking
statements, which speak only as of the date they are made. We undertake no
obligation to update or revise publicly any forward-looking statements,
whether because of new information, future events, or otherwise.

GENERAL

Embracing Our Future ...

Instruction fine-tuning

The text generation model can be instruction-tuned on any text data provided that the data is in the expected format. The instruction-tuned model can be further deployed for inference.

For input, use a training and optional validation directory. The train and validation directories should contain one or multiple JSON lines (.jsonl) formatted files. In particular, the train directory can also contain an optional *.json file describing the input and output formats.

The best model is selected according to the validation loss, calculated at the end of each epoch. If a validation set is not given, an (adjustable) percentage of the training data is automatically split and used for validation.

The training data must be formatted in a JSON lines (.jsonl) format, where each line is a dictionary representing a single data sample. All training data must be in a single folder; however, it can be saved in multiple .jsonl files. The .jsonl file extension is mandatory. The training folder can also contain a template.json file describing the input and output formats. If no template file is given, the following template will be used:

{
    "prompt": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{context}\n\n",
    "completion": "{response}"
}

In this case, the data in the JSON lines entries must include prompt and completion fields. If a custom template is provided, it must also use prompt and completion keys to define the input and output templates. The following is a sample custom template:

{
    "prompt": "question: {question} context: {context}",
    "completion": "{answer}"
}

Here, the data in the JSON lines entries must include the question, context, and answer fields.

The output is a trained model that can be deployed for inference.

We provide a subset of SEC filings data of Amazon. It is downloaded from publicly available EDGAR. For instructions on accessing the data, refer to Accessing EDGAR Data.

License: Creative Commons Attribution-ShareAlike License (CC BY-SA 4.0)

Related Updates

Recent Updates