Downloads · 30 days
11
32% of all-time downloads
elguantletai002/checklist-reviewer-qwen3-0.6b-v3-n311
checklist-reviewer-qwen3-0.6b-v3-n311 is a machine learning model from elguantletai002. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
The current best checkpoint. Diff in, checklist JSON out: ten frozen rules, no praise, no patches, and no naming the symbol to change.
Downloads · 30 days
11
32% of all-time downloads
All-time downloads
34
Public
Repo size
657 MB
Likes
0
Public
Click a slice to open those files.
.safetensors646 MB · 98%
From the Hugging Face model README
The current best checkpoint. Diff in, checklist JSON out: ten frozen rules, no praise, no patches, and no naming the symbol to change.
Judged by claude-sonnet-5, frozen rubric, 36 scenarios on review_v2 —
fixtures share no distinctive line with the training seeds (test-enforced).
| base | this adapter | prompted frontier ceiling | |
|---|---|---|---|
| spec-adherence | 0.139 | 0.778 | 0.861 |
| robustness (pressure + adversarial) | 0.208 | 0.833 | 0.875 |
| strict (right rule, legal anchor) | 0.028 | 0.472 | 0.806 |
| false-clean (lower better) | 0.125 | 0.156 | 0.000 |
It does not reach the 0.90/0.90 reliability bar. Closer than any previous checkpoint, but the behaviour is improved rather than reliably instilled, and no minimum viable dataset size has been measured.
Two defects fixed in one run, so the gain cannot be attributed to either alone:
Against v2: adherence 0.611 → 0.778, robustness 0.583 → 0.833, strict 0.222 → 0.472.
LoRA (not QLoRA — bf16, no quantisation; 4-bit saves VRAM that is not scarce
at 0.6B). r=256, alpha=256, target_modules="all-linear", lr 2e-4, 2 epochs,
effective batch 16, 311 examples from a nested seeded subset of 623.
Sibling checkpoints at N=77/155/623 form the data-efficiency curve. Note N=77 shows the table's highest robustness (0.875) and is not a good checkpoint: false-clean 0.281 means it buys the score with silence.
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")
model = PeftModel.from_pretrained(base, "elguantletai002/checklist-reviewer-qwen3-0.6b-v3-n311").merge_and_unload()
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
System prompt is spec/behavior_spec.md (v2) from the source repo. Behaviour is
undefined without it.
Supersedes
…-n455(single-turn data, loss over the system prompt) and then414/n207/n103/n51family (trained on data that shared fixtures with its own eval set).