Downloads · 30 days
5
10% of all-time downloads
MANGSEOK123/qwen3-4b-tau2-retail-posU132-1ep
qwen3-4b-tau2-retail-posU132-1ep is a machine learning model from MANGSEOK123. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Qwen3-4B-Instruct-2507 consolidated with OEL on the tau2-bench retail domain, using 132 externally authored task–memory pairs (recall 53 / composition 50 / substitution 29).
Downloads · 30 days
5
10% of all-time downloads
All-time downloads
52
Public
Parameters
4.4B
8.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors8.8 GB · 100%
From the Hugging Face model README
Qwen3-4B-Instruct-2507 consolidated with OEL on the tau2-bench retail domain, using 132 externally authored task–memory pairs (recall 53 / composition 50 / substitution 29).
| base model | Qwen3-4B-Instruct-2507 |
| training pairs | 132 |
| batch size | 12, every batch full |
| epochs | 1 |
| learning rate | 3e-6, constant |
| gradient clipping | 1.0 (verl default) |
| loss | full KL over all response tokens, kl_topk 256 |
| user simulator | gpt-4.1-mini (temperature 0) |
The student replays each task without memory; the teacher is the same weights with that task's memory in the system prompt. Only the prompt differs, and no reward is used.
| step | KL loss | entropy | grad norm |
|---|---|---|---|
| 1 | 0.006 | 0.203 | 1.119 |
| 2 | 0.008 | 0.194 | 0.841 |
| 3 | 0.007 | 0.246 | 4.245 |
| 4 | 0.013 | 0.357 | 1.397 |
| 5 | 0.012 | 0.413 | 0.795 |
| 6 | 0.006 | 0.227 | 0.587 |
| 7 | 0.014 | 0.187 | 1.909 |
| 8 | 0.011 | 0.337 | 2.907 |
| 9 | 0.023 | 0.032 | 4.325 |
| 10 | 0.015 | 0.291 | 2.053 |
| 11 | 0.027 | 0.207 | 1.204 |
Each step reads a different batch, so the loss column tracks batch difficulty rather than convergence.
Not evaluated. Pushed straight after training.
For context on the signal these memories carry: on a 190-pair retail training set the
base model scores avg 0.561 with no memory and 0.660 with each task's memory in the
prompt (+0.099, about 3.4 SE over 570 episodes). On the tau2-bench retail test split
the base model scores avg 0.400 / pass@3 0.575.
vllm serve MANGSEOK123/qwen3-4b-tau2-retail-posU132-1ep \
--enable-auto-tool-choice --tool-call-parser hermes --max-model-len 40960