GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zhengbang0707
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
zhengbang0707
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
Both refusal and safety judges: Qwen/Qwen3.5-27B (local).
Each prefix uses the same 128 safe and 128 unsafe examples, seed 42.
| Prefix | Safe refusal rate | Unsafe rate (safety judge) | Unsafe refusal rate | Failed judgments |
|---|---|---|---|---|
| no_prefix | 2.34% | 16.41% | 100.00% | 0 |
| sure_v1 | 3.12% | 16.41% | 100.00% | 0 |
| sure_v2 | 3.12% | 25.78% | 99.22% | 0 |
Rates use successfully judged answers; failed judgments are reported separately. Per-example outputs, all judge results and merged unsafe summaries are stored under each prefix directory.
Model: zhengbang0707/Qwen3.5-9B-eSFT2-KL-a0.1-lr5e-5-utility250-talos-20260925T011728Z; evaluated weight commit: 0645dcc85d8b64daa0292d6fd150d6bf6ae3c298.
| Benchmark | Samples | Metric | Score |
|---|---|---|---|
| ScienceQA-IMG test | 2,017 | Upstream exact match | 0.00% |
| VQAv2_val_lite | 500 | VQA soft accuracy | 0.00% |
Full fixed dataset splits; no random subsampling. Zero-shot, greedy decoding, thinking disabled, batch size 4, max_new_tokens 16, max_model_len 8192, BF16, SDPA, TP=2/DP=1 on two H100 NVL GPUs. The provided vision_utility_share prompts and scoring are preserved. The hardware guard was extended to accept H100; source hashes, pinned dataset revisions, processor checks, per-example outputs and runtime TP plans are included.
scienceqa_img: 2017/2017 answers contain only repeated <think> tags; 0 empty answers.vqav2_val_lite: 500/500 answers contain only repeated <think> tags; 0 empty answers.Selected examples reproduce this behavior with the trained checkpoint on a single GPU, batch size 1 or 4, BF16 or FP16, and both the training and utility environments. The original base model gives normal answers on the same diagnostic inputs. The reported full benchmark scores retain the original protocol; no answer parsing or generation setting was adjusted to avoid the repetitions. These small base-model checks are diagnostic controls, not a full baseline evaluation. See diagnostics/.