MiniCPM5-2B - OpenCodeReasoning-2 Python LoRA

QLoRA adapter for openbmb/MiniCPM5-2B, trained on a filtered subset of nvidia/OpenCodeReasoning-2.

Data (important)

This is not trained on the full corpus. OpenCodeReasoning-2 holds ~1.4M Python and ~1.1M C++ rows; this adapter uses 50000 rows (12053 unique questions), selected as:

  • split="python" only (C++ skipped)
  • judgement == "right" and pass_rate >= 0.8 (rows with pass_rate == -1 dropped)
  • unique-question first: the highest-pass_rate r1_generation per question_id, then extra high-pass solutions until 50000 rows
  • prompts reconstructed (OCR-2 stores question as "-") by joining question_id to nvidia/OpenCodeReasoning split_0 id -> input, with a dataset/split/index fallback into BAAI/TACO, codeparrot/apps, deepmind/code_contests and open-r1/codeforces
  • generations truncated to 12000 characters; very short prompts/generations dropped

OpenCodeReasoning-1 rows were not concatenated as extra training data - OCR-2 supersedes OCR-1 for generations, so mixing them would duplicate.

Training
Method 4-bit NF4 QLoRA (double quant, bf16 compute)
LoRA r=16, alpha=32, dropout=0.05
Targets q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainer TRL SFTTrainer, MiniCPM chat template
Schedule 1 epoch, seq 4096, bs 1 x grad-accum 8, lr 1e-4 cosine, warmup 0.03
Optimizer paged_adamw_8bit, gradient checkpointing
Hardware 1x A100 40GB (Modal)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("openbmb/MiniCPM5-2B", trust_remote_code=True, device_map="auto")
tok = AutoTokenizer.from_pretrained("NANI-Nithin/MiniCPM5-2B-OpenCodeReasoning2-LoRA", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "NANI-Nithin/MiniCPM5-2B-OpenCodeReasoning2-LoRA")

msgs = [{"role": "user", "content": "Write a Python function returning the n-th Fibonacci number."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=512)[0], skip_special_tokens=True))

Licensing and attribution

  • Adapter weights: Apache-2.0.
  • NVIDIA OpenCodeReasoning-2 reasoning traces: CC BY 4.0.
  • Original contest problem statements (TACO, APPS, CodeContests, Codeforces) keep their original licenses and terms; check each source dataset before redistributing.
  • Not affiliated with, endorsed by, or produced by OpenBMB or NVIDIA.

Citation

@article{ahmad2025opencodereasoning,
  title={OpenCodeReasoning: Advancing Data Distillation for Competitive Coding},
  author={Ahmad, Wasi Uddin and Narenthiran, Sean and Majumdar, Somshubra and Ficek, Aleksander and Jain, Siddhartha and Huang, Jocelyn and Noroozi, Vahid and Ginsburg, Boris},
  journal={arXiv preprint arXiv:2504.01943},
  year={2025}
}