Open-MOPD-SmolLM3-3B-RL-Math

Open-MOPD-SmolLM3-3B-RL-Math

This is the math-domain teacher in the Open-MOPD pipeline. It starts from BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-MixSFT and is trained only on math prompts with verifiable rewards using GRPO. This release corresponds to training step 100.

Training uses global batch size 128, mini-batch size 32, constant learning rate 1e-6 with 10 warmup steps, clipping at 0.2/0.25, rollout group size 16, temperature 1.0, a 30,000-token response limit, and no KL penalty. Groups with all-correct or all-incorrect generations are filtered, with up to eight resampling attempts.

Results

Math results use avg@64 with temperature 0.6. The broader evaluation setup uses max_model_len=32768, top_p=0.95, top_k=-1, and stop_token_ids=[128012].

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-RL-Math"
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 math 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