Code-13B
Large Language Models (LLMs) are good with code generations. Sometimes they do make mistakes in code generation. How about if they can give detailed explanation along with the code. This is what I have tried over here. The base Llama-2 model was used for training purpose. It is trained on around 74000 set of codes. Each set having 2 conversations. Along with Python, Java, JavaScript, GO, C++, Rust etc. code with detailed explanation is used for training purpose. It is built upon using my existing Dataset Python-Code-23k-ShareGPT. This conversation is in Vicuna/ShareGPT format. Each set, along with code, has detailed explanation.
I have released the new data Code-74k-ShareGPT on which this Model is trained.
Training:
Entire dataset was trained on Azure 4 x A100 80GB. For 3 epoch, training took 42 hours. DeepSpeed codebase was used for training purpose. This was trained on Llama-2 by Meta.
This is a full fine tuned model. Links for quantized models are given below.
GPTQ GGUF & AWQ
GPTQ: Link
GGUF: Link
AWQ: Link
Extremely thankful to TheBloke for making Quantized versions of model.
Example Prompt:
This is a conversation with your helpful AI assistant. AI assistant can generate Code in various Programming Languages along with necessary explanation.
Context
You are a helpful AI assistant.
USER: <prompt>
ASSISTANT:
You can modify above Prompt as per your requirement. I have used ShareGPT/Vicuna format v1.1 .
I want to say special Thanks to the Open Source community for helping & guiding me to better understand the AI/Model development.
Thank you for your love & support.
Example Output
- Navier-Stokes Equation Solver
- KSC Complexity
- GO
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
Metric | Value |
---|---|
Avg. | 54.81 |
AI2 Reasoning Challenge (25-Shot) | 57.34 |
HellaSwag (10-Shot) | 83.28 |
MMLU (5-Shot) | 53.17 |
TruthfulQA (0-shot) | 42.46 |
Winogrande (5-shot) | 73.56 |
GSM8k (5-shot) | 19.03 |
- Downloads last month
- 1,315
Model tree for ajibawa-2023/Code-13B
Dataset used to train ajibawa-2023/Code-13B
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard57.340
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard83.280
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard53.170
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard42.460
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard73.560
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard19.030