Downloads · 30 days
0
aprm-sft-tau2-retail
mzio/aprm-sft-tau2-retail
aprm-sft-tau2-retail is a machine learning model from mzio. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
LoRA adapters (r8a16, base Qwen/Qwen3-4B-Instruct-2507) from Act-PRM supervised fine-tuning on tau2-bench retail. Variants: actionsonly (baseline), expertthoughts (oracle), thoughts{policy,base}[last] (Act-PRM inferre…
Access
Public
Updated Aug 31, 2026
Repo size
2.8 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors1.8 GB · 100%
At a glance
- Access
- Public
- Created
- Jul 28, 2026
- Updated
- Aug 31, 2026
- SHA
- c016df4d
Model card
From the Hugging Face model README
Act-PRM SFT LoRA checkpoints — tau2 retail
LoRA adapters (r8_a16, base Qwen/Qwen3-4B-Instruct-2507) from Act-PRM supervised fine-tuning on tau2-bench retail. Variants: actions_only (baseline), expert_thoughts (oracle), thoughts_{policy,base}[_last] (Act-PRM inferred thoughts), each in hide-observations and full-context regimes. Adapters live under retail/<variant_regime>/.
Held-out action-only eval (lower PPL / higher acc = better next-action fit)
| variant_regime | action-only PPL | action-acc |
|---|---|---|
| expert_thoughts_lr1e_3_adamw_nb100_heldout | 1.663 | 0.846 |
| expert_thoughts_lr1e_3_adamw_nb100_heldout | 1.663 | 0.846 |
See the project for methodology (Act-PRM: infer latent thoughts behind action-only demos via offline EM).
- Created
- Jul 28, 2026
- Updated
- Aug 31, 2026