Qwen3.5-4B-MBT-R / README.md
12kimih's picture
Cite the arXiv entry of the paper
a80aea7
|
Raw History Blame Contribute Delete
2.94 kB
---
license: apache-2.0
base_model: Qwen/Qwen3.5-4B-Base
base_model_relation: finetune
datasets:
- 12kimih/mbt-traces
language:
- en
library_name: transformers
pipeline_tag: text-generation
tags:
- reasoning
- math
- metacognition
- sft
---
# Qwen3.5-4B-MBT-R
[Qwen/Qwen3.5-4B-Base](https://huggingface.co/Qwen/Qwen3.5-4B-Base) fine-tuned on traces edited by MBT-R: the editor rewrites the whole trace in one call.
It accompanies the paper **"Distilling Metacognitive Behaviors through Agentic Editing"**.
| Model | Training data |
|---|---|
| [Qwen3.5-4B-Unedited-SFT](https://huggingface.co/12kimih/Qwen3.5-4B-Unedited-SFT) | Unedited teacher traces |
| [Qwen3.5-4B-MBT-R](https://huggingface.co/12kimih/Qwen3.5-4B-MBT-R) | MBT-R (rewrite) |
| [Qwen3.5-4B-MBT-I](https://huggingface.co/12kimih/Qwen3.5-4B-MBT-I) | MBT-I (insert) |
| [Qwen3.5-4B-MBT-A](https://huggingface.co/12kimih/Qwen3.5-4B-MBT-A) | MBT-A (agentic) |
Training data: built from the traces in
[`12kimih/mbt-traces`](https://huggingface.co/datasets/12kimih/mbt-traces).
Code and prompts:
[GitHub](https://github.com/12kimih/metacognitive-behavioral-tuning).
## Usage
The chat template that ships with the model opens a `<think>` block, so the generated text is the
reasoning, then `</think>` and the answer. End the question with the instruction used in training:
```python
from vllm import LLM, SamplingParams
llm = LLM("12kimih/Qwen3.5-4B-MBT-R")
question = "Find the remainder when 2^100 is divided by 7."
prompt = question + "\n\nPlease reason step by step, and put your final answer within \\boxed{}."
params = SamplingParams(temperature=1.0, top_p=0.95, max_tokens=122880)
out = llm.chat([{"role": "user", "content": prompt}], params)
print(out[0].outputs[0].text)
```
The weights are bf16. Long answers are common, so leave room for tens of thousands of tokens.
## Training
| Setting | Value |
|---|---|
| Base model | Qwen/Qwen3.5-4B-Base |
| Method | Full-parameter supervised fine-tuning, loss on assistant tokens only |
| Epochs | 15 |
| Optimizer | Muon (Megatron implementation), learning rate 1e-5, weight decay 0 |
| Schedule | Constant with 40 warmup steps |
| Global batch | 64 sequences (8 GPUs x 8 accumulation steps), padding-free packing |
| Maximum length | 131,072 tokens |
| Precision | bf16 autocast with fp32 master weights; released in bf16 |
## License
Apache-2.0, following the base model. The training data is CC BY-SA 4.0; see the dataset card for
upstream attribution.
## Citation
```bibtex
@misc{kim2026distilling,
title = {Distilling Metacognitive Behaviors through Agentic Editing},
author = {Kim, Ik-hwan and Han, Hyeongrok and Jung, Mingi and Yu, Sangwon and Hong, Jinseok and Kim, Sang Hun and Choi, Yoonyoung and Yoon, Sungroh},
year = {2026},
eprint = {2602.22508},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2602.22508}
}
```