Downloads · 30 days
8
15% of all-time downloads
MANGSEOK123/qwen3-4b-tau2-retail-fp105-1ep
qwen3-4b-tau2-retail-fp105-1ep 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 105 externally authored task–memory pairs (recall 54 / composition 31 / substitution 20). The recall share here is about half the set…
Downloads · 30 days
8
15% 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 105 externally authored task–memory pairs (recall 54 / composition 31 / substitution 20). The recall share here is about half the set, noticeably higher than in the sibling runs.
Siblings: fp190, fp300, fp117.
| base model | Qwen3-4B-Instruct-2507 |
| training pairs | 105 |
| batch / steps | 12 / 9 (1 epoch, last batch 9) |
| 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 | 0.010 | 0.294 | 1.205 |
| 2 | 0.011 | 0.258 | 0.978 |
| 3 | 0.009 | 0.335 | 2.300 |
| 4 | 0.013 | 0.406 | 1.015 |
| 5 | 0.013 | 0.301 | 1.415 |
| 6 | 0.007 | 0.255 | 0.847 |
| 7 | 0.008 | 0.202 | 1.214 |
| 8 | 0.010 | 0.289 | 1.053 |
| 9 | 0.005 | 0.328 | 2.618 |
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 what the teacher can do: on the 190-pair training set, the base model scores 0.561 without memory and 0.660 with each task's memory in the prompt (+0.099, about 3.4 SE over 570 episodes). The memories carry real signal; the open question in this series is how little of it the distillation step transfers.
vllm serve MANGSEOK123/qwen3-4b-tau2-retail-fp105-1ep \
--enable-auto-tool-choice --tool-call-parser hermes --max-model-len 40960