Modelpublic

mrzhao13/qwen3.5-2b-shopsimulator-sft-512-1ep

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes15downloads

Model Card

Qwen3.5-2B ShopSimulator SFT-512 1 Epoch

This is a Hugging Face export of Qwen/Qwen3.5-2B after one epoch of supervised fine-tuning on successful ShopSimulator teacher trajectories. Training used the Pi agent harness and the modified Slime integration.

The fine-tuning data is text-only. The checkpoint retains the original multimodal architecture, but visual capabilities were not trained or evaluated in this experiment.

Training

Item	Value
Base model	`Qwen/Qwen3.5-2B`
Teacher	`deepseek-v4-flash`, thinking disabled
Collection	512 tasks, one candidate per task
Accepted trajectories	412 (80.47% task coverage)
Turn-level SFT examples	6,153
Epochs	1
Optimizer steps	2,051
Global batch size	3
Learning rate	`1e-5`
Maximum tokens per GPU	12,288
Longest prepared example	8,162 tokens at conversion time
Loss mask	Qwen3.5 multi-turn mask; template-injected empty think blocks excluded
Training framework	Slime + Megatron-LM
Hardware	One NVIDIA Pro 6000D 84 GB
Weight identity SHA-256	`8a98272246a3dbd4c59d64e3981807e6b7719007880cf2aed9fe0996e76f8a7d`

The training dataset repository is currently private while the redistribution status of the upstream ShopSimulator task content is being clarified.

Evaluation

The following results use the same modified ShopSimulator service, deterministic price generation, fixed official_test_200 task slice, and one rollout per task. These are k=1 point estimates, not uncertainty estimates.

Model	Positive-reward pass@1	Strict-success pass@1	mean@1 `r_loose`	mean@1 `r_hard`
Qwen3.5-2B	2.0%	0.0%	0.004286	0.000000
This SFT checkpoint	72.5%	10.5%	0.389829	0.124417

Usage

Use this repository as a drop-in replacement for Qwen/Qwen3.5-2B with a recent Transformers version that supports Qwen3.5. Refer to the official Qwen3.5-2B model card for loading and inference examples, replacing the model ID with this repository.

For ShopSimulator agent evaluation, use the Pi extension, prompt template and environment adapter from the companion pi-slime-shopsimulator source repository when it is published.

Limitations

—Fine-tuned only on Chinese ShopSimulator trajectories.
—Evaluated only in the patched text-based ShopSimulator environment.
—The deterministic pricing patch changes scores relative to the unpatched upstream environment.
—The official evaluation contains 200 tasks with one stochastic rollout each.
—Visual behavior, general instruction following and deployment safety were not evaluated.
—Teacher-generated content may contain errors.

License and attribution

The base model and these derivative weights are distributed under Apache-2.0. The original Qwen license is included in LICENSE. This model card identifies the checkpoint as modified and does not grant rights to third-party ShopSimulator task or environment content.