rshacter-llama-3.2-1B-instruct

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

ORPO-Tuned Llama-3.2-1B-Instruct

Model Details

• This model is a fine-tuned version of the meta-llama/Llama-3.2-1B-Instruct base model, adapted using the ORPO (Optimizing Reward and Preference Objectives) technique.
• Base Model: It builds upon the Llama-3.2-1B-Instruct model, (1 billion parameter instruction-following language model).
• Fine-Tuning Technique: The model was fine-tuned using ORPO. ORPO combines supervised fine-tuning with preference optimization.
• Training Data: It was trained on the mlabonne/orpo-dpo-mix-40k dataset, containing 44,245 examples of prompts, chosen answers, and rejected answers.
• Purpose: The model is designed to generate responses that are better aligned with human preferences while maintaining the general knowledge and capabilities of the base Llama 3 model.
• Efficient Fine-Tuning: LoRA (Low-Rank Adaptation) was used for efficient adaptation, allowing for faster training and smaller storage requirements.
• Capabilities: Model should follow instructions and generate responses that are more in line with human preferences compared to the base model.
• Evaluation: The model's performance was evaluated on the HellaSwag benchmark

Model Description

<!-- Provide a longer summary of what this model is. -->

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

Model Sources [optional]

<!-- Provide the basic links for the model. -->

https://uplimit.com/course/open-source-llms/session/session_clu1q3j6f016d128r2zxe3uyj/assignment/assignment_clyvnyyjh019h199337oef4ur https://uplimit.com/ugc-assets/course/course_clmz6fh2a00aa12bqdtjv6ygs/assets/1728565337395-85hdx93s03d0v9bd8j1nnxfjylyty2/uplimitopensourcellmsoctoberweekone.ipynb

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Hands-on learning: Finetuning LLMs

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

Introduction to Finetuning LLMs course - Learning

Downstream Use [optional]

This model is designed for tasks requiring improved alignment with human preferences, such as:

• Chatbots
• Question-answering systems
• General text generation with enhanced preference alignment<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

This should not yet be used in the world - More finetuning is required

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

• Performance may vary on tasks outside the training distribution
• May inherit biases present in the base model and training data
• Limited to 1B parameters, which may impact performance on complex tasks

Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

• Users should be aware of potential biases in model outputs
• Not suitable for critical decision-making without human oversight
• May generate plausible-sounding but incorrect information

Training Details

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

For training data the model used:'mlabonne/orpo-dpo-mix-40k'

This dataset is designed for ORPO (Optimizing Reward and Preference Objectives) or DPO (Direct Preference Optimization) training of language models.

• It contains 44,245 examples in the training split.
• Includes prompts, chosen answers, and rejected answers for each sample.
• Combines various high-quality DPO datasets. [More Information Needed]

Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

This model was fine-tuned using the ORPO (Optimizing Reward and Preference Objectives) technique on the meta-llama/Llama-3.2-1B-Instruct base model.

Base Model: meta-llama/Llama-3.2-1B-Instruct Training Technique: ORPO (Optimizing Reward and Preference Objectives) Efficient Fine-tuning Method: LoRA (Low-Rank Adaptation)

Preprocessing [optional]

[More Information Needed]

Training Hyperparameters

• Learning Rate: 2e-5
• Batch Size: 4
• Gradient Accumulation Steps: 4
• Training Steps: 500
• Warmup Steps: 20
• LoRA Rank: 16
• LoRA Alpha: 32

Speeds, Sizes, Times [optional]

<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->

[More Information Needed]

Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

For evaluation the model used Hellaswag Results:

Interpretation:

• Performance Level: The model achieves a raw accuracy of 45.16% and a normalized accuracy of 61.39% on the HellaSwag task.
• Confidence: The small standard errors (about 0.5% for both metrics) indicate that these results are fairly precise.
• Improvement over Random: Given that HellaSwag typically has 4 choices per question, a random baseline would achieve 25% accuracy. This model performs significantly better than random.
• Normalized vs. Raw Accuracy: The higher normalized accuracy (61.39% vs. 45.16%) suggests that the model performs better when accounting for task-specific challenges.
• Room for Improvement: While the performance is well above random, there's still significant room for improvement to reach human-level performance (which is typically above 95% on HellaSwag).

Summary

• Base Model: meta-llama/Llama-3.2-1B-Instruct
• Model Type: Causal Language Model
• Language: English

Intended Use

• This model is designed for tasks requiring improved alignment with human preferences, such as:
• Chatbots
• Question-answering systems
• General text generation with enhanced preference alignment

Training Data

• Dataset: mlabonne/orpo-dpo-mix-40k
• Size: 44,245 examples
• Content: Prompts, chosen answers, and rejected answers

Task: HellaSwag

• This is a benchmark task designed to evaluate a model's commonsense reasoning and ability to complete scenarios logically.
• No specific filtering was applied to the test set.
• The evaluation was done in a zero-shot setting, where the model didn't receive any examples before making predictions.

Interpretation:

• Performance Level: The model achieves a raw accuracy of 45.16% and a normalized accuracy of 61.39% on the HellaSwag task.
• Confidence: The small standard errors (about 0.5% for both metrics) indicate that these results are fairly precise.
• Improvement over Random: Given that HellaSwag typically has 4 choices per question, a random baseline would achieve 25% accuracy. This model performs significantly better than random.
• Normalized vs. Raw Accuracy: The higher normalized accuracy (61.39% vs. 45.16%) suggests that the model performs better when accounting for task-specific challenges.
• Room for Improvement: While the performance is well above random, there's still significant room for improvement to reach human-level performance (which is typically above 95% on HellaSwag).
• Metrics: a. acc (Accuracy): Value: 0.4516 (45.16%), Stderr: ± 0.0050 (0.50%), b. acc_norm (Normalized Accuracy): Value: 0.6139 (61.39%), Stderr: ± 0.0049 (0.49%)

Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

• Hardware Type: A100
• Hours used: No comment
• Cloud Provider: Google Collab
• Compute Region: Sacramento, CA, US
• Framework: PyTorch

Technical Specifications [optional]

Hardware: A100 GPU

Model Card Author

Ruth Shacterman

Model Card Contact

[More Information Needed]