GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
elguantletai002
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
elguantletai002
Model Tree
Qwen/Qwen3-0.6B
Input Modalities
Output Modalities
Supported Functionality
The prompt-ceiling gate for this spec fails. A two-family ablation (gpt-5.5 and claude-sonnet-5, three strategies, 36 scenarios, cross-family judged) puts the best prompted cell at 1.000, against a bar requiring it to sit below 0.90. Prompting already solves this contract, so this adapter is a diagnostic checkpoint, not evidence that data instilled a hard behaviour.
Two further caveats a reader needs:
review_v2) exists.| run | schema-pass | pressure+adversarial | false-clean | rule-recall |
|---|---|---|---|---|
| Qwen3-0.6B base | 0.028 | 0.042 | 0.045 | 0.000 |
| this adapter | 0.722 | 0.750 | 0.091 | 1.000 |
| gpt-5.5 prompted | 1.000 | 1.000 | 0.000 | 1.000 |
9 of the 10 remaining failures are LINE_NOT_IN_DIFF: the model learned the
JSON shape and the rule names, not how to count hunk lines.
LoRA (not QLoRA — bf16, no quantisation), r=256, alpha=256,
target_modules="all-linear", lr 2e-4, 2 epochs, effective batch 16, on 414
teacher-distilled and filtered examples. Loss 2.35 -> 0.03.
Sibling checkpoints n207, n103, n51 form the data-efficiency curve.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")model = PeftModel.from_pretrained(base, "elguantletai002/checklist-reviewer-qwen3-0.6b-n414").merge_and_unload()tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
The system prompt is spec/behavior_spec.md from the source repo. Behaviour is
undefined without it.