---
license: apache-2.0
library_name: transformers
base_model: Qwen/Qwen3-8B-Base
pipeline_tag: text-generation
tags:
- qwen3
- on-policy-distillation
- multi-teacher
- reinforcement-learning
---

# mopd-iter200

Multi-teacher **On-Policy Distillation (OPD)** checkpoint of a Qwen3-8B student,
exported at **iteration 200**.

## Overview

- **Architecture:** Qwen3-8B (dense, 36 layers, hidden 4096, GQA 32Q/8KV, vocab 151936, 32k context)
- **Training:** on-policy distillation where a single student rolls out a mixed
  math + search + tool-use (tau) stream; the only training signal is per-token
  reverse-KL to a domain-specific teacher (static domain routing, task reward = 0).
- **Init:** Qwen3-8B SFT chain (Math -> Sea-SFT -> Search -> Tau-SFT -> IF).
- **Format:** converted from a Megatron torch_dist checkpoint to HuggingFace
  safetensors (bf16).

## Usage

\`\`\`python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "willamazon1/mopd-iter200"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "What is 12*8?"}]
text = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
ids = tok(text, return_tensors="pt").input_ids.to(model.device)
out = model.generate(ids, max_new_tokens=64)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
\`\`\`