qwen25-coder-32b-sft-ocr2-combined

Qwen2.5-Coder-32B — SFT on OpenCodeReasoning-2

Fine-tuned from Qwen/Qwen2.5-Coder-32B-Instruct using supervised fine-tuning (SFT) on nvidia/OpenCodeReasoning-2 — a dataset of competitive programming problems with long chain-of-thought reasoning traces verified by code execution.

Training details

• Base model: Qwen/Qwen2.5-Coder-32B-Instruct
• Dataset: nvidia/OpenCodeReasoning-2 (TACO + APPS + CodeContests + Codeforces, all sources combined)
• Method: SFT with QLoRA (4-bit NF4, r=64, alpha=128), merged back to full precision
• Max sequence length: 32 768 tokens
• Hardware: A100 SXM4 80 GB

Use

The model follows the Qwen2.5-Coder chat template. Given a competitive programming problem, it produces a long chain-of-thought reasoning trace followed by a Python or C++ solution.

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Belaleatsbanana/qwen25-coder-32b-sft-ocr2-combined",
    device_map="auto",
    load_in_4bit=True,
)
tokenizer = AutoTokenizer.from_pretrained("Belaleatsbanana/qwen25-coder-32b-sft-ocr2-combined")