metadata
base_model: unsloth/Qwen2.5-Coder-7B-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- qwen2
license: apache-2.0
language:
- en
datasets:
- open-r1/codeforces-cots
Qwen2.5-Coder-7B-Codeforces
This model is a fine-tuned version of Qwen2.5-Coder-7B (quantized in 4-bit via QLoRA). It has been specifically trained to act as an intelligent programming tutor and expert solver for Codeforces competitive programming problems.
It is designed to serve as the generation node in a broader RAG (Retrieval-Augmented Generation) architecture, dynamically adapting its response based on the structured instruction provided in the prompt.
Key Features
- Dual-Mode Inference: Can switch between generating a progressive, 1-2 sentence theoretical hint (Tutor Mode) or a fully functional, optimized Python solution (Expert Mode).
- Memory Efficient: Fine-tuned using Unsloth and optimized with an 8-bit Paged AdamW optimizer to compress a 7B model workflow into a single 16GB T4 GPU envelope.
- Context Preservation: Maintained a robust 2048/3072 sequence length to handle complex problem statements and retrieved vector database context without dropping long-dependency tokens.
Prompt Template
To get the exact structured output and prevent hallucinations, you must use the following prompt format when querying the model:
1. Tutor Mode (For Hints Only)
Instruction: You are a programming tutor. Give ONE short hint for this problem. Do NOT give code or reveal the full solution. Just the key insight in 1-2 sentences.
Difficulty Rating: [e.g., 1300]
Topics: [e.g., greedy, math, sortings]
Problem:
[Insert Codeforces Problem Text Here]
Hint: