Shaurya-saini/qwen2.5-coder-7b-apps-qlora
Shaurya-saini/qwen2.5-coder-7b-apps-qlora is a 7.6 billion parameter model, fine-tuned from Qwen2.5-Coder-7B-Instruct using QLoRA on the APPS competitive programming dataset. This model was developed by Shaurya-saini as a study on performance changes after fine-tuning on specific code challenges. It is intended for Python competitive programming problems, though it currently underperforms its base model across all difficulty tiers. The model's primary purpose is for transparency and reproducibility in code model fine-tuning experiments.
Loading preview...
Model Overview
Shaurya-saini/qwen2.5-coder-7b-apps-qlora is a 7.6 billion parameter model derived from Qwen2.5-Coder-7B-Instruct. It was fine-tuned using QLoRA (4-bit base, LoRA rank 16) with Unsloth, specifically on the APPS competitive programming training split. The training involved one epoch over approximately 620 steps, aiming to adapt the model for Python competitive programming tasks.
Key Characteristics and Performance
This model is published as a portfolio study to demonstrate before/after performance on execution-based code benchmarks. Crucially, in evaluations on the held-out APPS test split, this fine-tuned model underperforms its base model (Qwen2.5-Coder-7B-Instruct) across all difficulty tiers (Easy, Medium, Hard). For instance, on Medium difficulty problems, the base model achieved 7.3% pass@1, while this fine-tune achieved 2.0%.
Intended Use and Limitations
- Intended Use: Designed for Python competitive-programming style problems, requiring the Qwen chat template for prompting.
- Limitations: The observed performance regression is attributed to several factors:
- Training data selection: Using short, golfed solutions from APPS led to the model imitating terse, cruder code.
- Catastrophic forgetting: One epoch on a narrow dataset eroded the base model's broader coding abilities.
- Learned syntax artifact: The model frequently over-produces closing brackets, causing compilation errors.
This model is provided for transparency and reproducibility of the experimental results, rather than as an improved solution over its base model. Future iterations aim to address these limitations.