Open-MOPD-SmolLM3-3B-RL-Code
This is the code-domain teacher in the Open-MOPD pipeline. It starts from
BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-MixSFT and is trained only on code
prompts with verifiable rewards using GRPO. This release corresponds to
training step 180.
Training uses global batch size 128, mini-batch size 32, learning rate 1e-6,
rollout group size 16, a 30,000-token response limit, no KL penalty, and
accuracy-based group filtering.
Results
| Model | LiveCodeBench v5 | LiveCodeBench v6 | Code average |
|---|---|---|---|
| RL-Code teacher | 22.16 | 21.31 | 21.73 |
| MixSFT starting point | 15.99 | 19.20 | 17.60 |
Results use avg@10 rather than best@10, with temperature 1.0,
max_model_len=32768, top_p=0.95, top_k=-1, and
stop_token_ids=[128012].
The code portion of the RL prompt mixture explicitly excludes LiveCodeBench.
The decontamination record is available in
BytedTsinghua-SIA/Open-MOPD-Data under rl_prompt_mix/manifest.json.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-RL-Code"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")
Intended use and limitations
This is a domain teacher intended for distillation, not a general-purpose assistant. It was optimized only on code and can perform worse than MixSFT on other domains.
Model specifications
- Architecture:
SmolLM3ForCausalLM - Parameters: approximately 3B
- Layers: 36
- Vocabulary size: 128,256
- Weights: BF16, approximately 6.2 GB
- Includes tokenizer and chat template