GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Jazhyc
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Jazhyc
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
Rank-16/alpha-32 LoRA for Qwen/Qwen3.5-9B, trained for continuous
direct-boundary deception scoring in Aletheia's Quest.
The teacher was moonshotai/kimi-k3, served by Fireworks through OpenRouter.
It scored normalized literal 0|1 log probabilities immediately after
Prediction: with the frozen no-thinking Truth Value Guard prompt. The student
used all 2,880 varied-deception training rows for two epochs with AdamW at
5e-5, effective batch size 32, and binary soft-target BCE. It received no
generated teacher reasoning, hard-label loss, completion loss, or pairwise
loss.
The frozen Qwen/Qwen3.5-9B base ran in BF16. Soft targets, LoRA weights,
AdamW state, binary margins, and BCE were FP32. The 256 adapter tensors use
canonical Qwen3.5 paths under model.language_model.layers and explicitly
exclude visual modules.
Frozen local-validation direct-margin results:
| metric | value |
|---|---|
| Macro AUROC | 0.95994 |
| Instructed AUROC | 0.99833 |
| Varied AUROC | 0.90875 |
| Balanced accuracy at 0.5 | 0.91548 |
These are local competition diagnostics and do not guarantee private-dataset
transfer. The adapter is intended for continuous literal 0|1 boundary
scoring; it does not generate judge reasoning at inference.