prvnsmpth's picture
Update README.md
57fffcd verified
metadata
library_name: transformers
tags:
  - unsloth
  - fim
  - code-completion
license: mit
base_model:
  - unsloth/Qwen2.5-Coder-14B-bnb-4bit

qwen2.5-coder-14b-finetune

This model is a fine-tune of Qwen 2.5 Coder 14B on code from a single repo.

More details: https://prvn.sh/build-your-own-github-copilot/

How to Get Started with the Model

See Jupyter notebook here: https://github.com/prvnsmpth/finetune-code-assistant

Training Details

Training Data

Training data was generated from source files in https://github.com/hcengineering/platform. Dataset generation script: https://github.com/prvnsmpth/finetune-code-assistant/tree/master/src/dataset_gen

Evaluation

The model was tested on a 2% holdout of source files in the target repository.

Metrics

  1. Exact match
  2. Corpus BLEU score

Summary

The fine-tuned model shows a 47% uplift in exact match FIM performance versus the baseline.