Open-MOPD-SmolLM3-3B-MixSFT

Open-MOPD-SmolLM3-3B-MixSFT

This model is the mixed-domain supervised fine-tuning initialization used by the Open-MOPD pipeline. All three domain-specific RL teachers and the final multi-teacher distilled student start from this checkpoint.

The model is derived from HuggingFaceTB/SmolLM3-3B-Base through four epochs of supervised fine-tuning on math, code, and instruction-following data. The training setup uses global batch size 128, learning rate 4e-5, cosine decay with 3% warmup, a maximum sequence length of 32,768, and 30,116 optimization steps.

Domains are balanced by response-token count rather than example count. This prevents the 820K shorter instruction-following responses from overwhelming the smaller but longer math and code corpora. After balancing, math, code, and instruction following contribute approximately 37.3%, 28.1%, and 34.6% of training response tokens.

Results

Evaluation protocol

• Math: AIME24 and AIME25, avg@64, temperature 0.6.
• Code: LiveCodeBench v5 and v6, avg@10, temperature 1.0.
• Instruction following: IFEval and IFBench_test, n=1, temperature 1.0, with enable_thinking=true.

All evaluations use max_model_len=32768, top_p=0.95, top_k=-1, and stop_token_ids=[128012]. Scores are averaged per dataset, then per domain, followed by a macro-average across domains.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-MixSFT"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="bfloat16",
    device_map="auto",
)

messages = [{"role": "user", "content": "Write a Python function that merges two sorted lists."}]
inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=True,
).to(model.device)
outputs = model.generate(inputs, max_new_tokens=4096, temperature=0.6, top_p=0.95, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Role in the pipeline

SmolLM3-3B-Base -> MixSFT initialization -> Math/Code/IF RL teachers -> Open-MOPD Final.

Model specifications

• Architecture: SmolLM3ForCausalLM
• Parameters: approximately 3B
• Layers: 36
• Vocabulary size: 128,256
• Weights: BF16, approximately 6.2 GB
• Includes tokenizer and chat template

The published tokenizer metadata is compatible with Transformers 4.x and produces the same tokenization and chat-template sequences as the original training artifact.