wisdompan/qwen35-2b-mopd-mixsft

Qwen3.5-2B MOPD MixSFT

This repository contains the full-parameter MixSFT initialization used for an Open-MOPD reproduction based on Qwen3.5-2B. It is intended as the common starting point for the Math, Code, and Instruction-Following teacher policies.

Status: MixSFT checkpoint, not the final MOPD student. Teacher RL and student distillation checkpoints will be published separately after they pass stability checks and evaluation.

基于 Qwen3.5-2B 的 MOPD 复现实验 MixSFT 初始化权重，覆盖数学、 代码和指令遵循数据。它不是最终 MOPD student；后续稳定的 teacher 和 student 权重会分阶段发布。

Model details

The checkpoint was trained with Adafactor, a cosine learning-rate schedule, peak learning rate 1e-5, weight decay 0.1, gradient accumulation 8, gradient checkpointing, and seed 20260902. The final recorded training loss was 0.6251 and token accuracy was 0.8485.

Intended use

• —Research on multi-domain reasoning and MOPD-style teacher/student training.
• —Continued Math, Code, and Instruction-Following RL training.
• —Local inference and controlled evaluation of this intermediate SFT stage.

This model is not presented as a production-ready assistant or as the final result of the MOPD experiment.

Quick start

Qwen3.5 requires a recent Transformers build. The serving flow follows the upstream Qwen3.5 recommendation:

pip install "transformers[serving] @ git+https://github.com/huggingface/transformers.git@main"
transformers serve \
  --force-model wisdompan/qwen35-2b-mopd-mixsft \
  --port 8000 \
  --continuous-batching

Then call the OpenAI-compatible endpoint:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused")
response = client.chat.completions.create(
    model="wisdompan/qwen35-2b-mopd-mixsft",
    messages=[{"role": "user", "content": "Solve: 2x + 3 = 11"}],
)
print(response.choices[0].message.content)

Evaluation status

Only a six-example pipeline sanity check was bundled with this checkpoint. It is useful for confirming that generation and reward routing work, but it is too small to report as a benchmark. Broader evaluation will accompany the final teacher and student releases.

Limitations

• —This is an intermediate checkpoint and may produce incorrect, verbose, or unsafe content.
• —Math, coding, and instruction-following quality has not yet been established on broad public benchmarks.
• —Users should independently evaluate safety and task fitness before deployment.

Training lineage

Qwen/Qwen3.5-2B
  -> MixSFT (this repository)
  -> Math / Code / IF teacher RL
  -> MOPD student distillation

License

This derivative is released under the Apache License 2.0, following the base model license. Users must also comply with the terms and usage requirements of the upstream Qwen3.5-2B model.