Downloads · 30 days
13
31% of all-time downloads
xiaorui638/Qwen3.5-9B-ZwZ-S57-verify-anneal
Qwen3.5-9B-ZwZ-S57-verify-anneal is a image-text-to-text model from xiaorui638. Use it for the image-text-to-text 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.
GRPO-trained on 57,447 false-premise twin pairs (positive/negative, 2:1), where each negative asks about something not in the image and the model must pick a premise-denial option instead of answering.
Downloads · 30 days
13
31% of all-time downloads
All-time downloads
42
Public
Parameters
9.4B
18.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors18.8 GB · 100%
From the Hugging Face model README
GRPO-trained on 57,447 false-premise twin pairs (positive/negative, 2:1), where each negative asks about something not in the image and the model must pick a premise-denial option instead of answering.
Reward is gated_mult: acc * (alpha + (1-alpha) * coverage), where coverage
is how many gold sub-questions the model's own emitted <check> questions
cover, scored by an LLM judge. The curriculum retires that scaffold:
alpha 0.5→1.0 and the format weight 0.1→0 linearly over steps 40–80, so the
last ~70 steps are pure outcome-only GRPO from a verify-shaped policy.
Motivation: coverage saturates (~0.95 by step 60), after which the scaffold adds no gradient but keeps rehearsing a ~75%-yes verdict distribution that degrades out-of-domain denial calibration. Annealing it away fixed that.
| metric | base | verify (no anneal) | this model |
|---|---|---|---|
| ZTB-mcq paired (in-domain) | 19.04 | 50.03 | 53.55 |
| DASH-B acc_no | 69.43 | 52.35 | 69.13 |
| DASH-B overall | 68.53 | 77.26 | 82.14 |
| VERVE paired | 66.79 | 79.13 | 81.65 |
53.55 paired is the best in-domain result in the project. The un-annealed arm's
OOD yes-bias (DASH-B acc_no 52.35) is largely repaired.
Use the verify v2.1 prompt it was trained with — under the default prompt
in-domain paired drops to ~42.6. Needs vLLM >= 0.18 (model_type=qwen3_5);
weights are already in the stock HF key layout (model.visual.*).
Single seed. Trained for in-domain false-premise robustness: general-ability
aggregates are at or slightly below the plain-GRPO arm (mcq-mean 70.97 vs
71.68), and mme-realworld-lite carries a persistent ~4-point tax vs base.