Shield-Qwen3Guard-Gen-8B-FT-PEFT-CE

Full fine-tune of Qwen/Qwen3Guard-Gen-8B on the DIA-GUARD train split (~836K samples, 50 English dialects). Serves as the 8B teacher in the Shield-*-KD-*-Qwen3Guard-Gen-8B-FT-* knowledge-distillation family in this org.

Provenance

This repo re-hosts the full-parameter fine-tune produced by @aliwister in aliwister/qwen3guard_gen_8b-diaguard. It replaces an earlier LoRA-CE attempt by us which collapsed to a trivial predictor; aliwister's full FT substantially outperforms that attempt (see table below) so we standardise on this checkpoint as the canonical Shield 8B teacher.

Note on the repo name: the -FT-PEFT-CE suffix is inherited from our earlier LoRA-CE checkpoint name so that the existing KD child repos (jsl5710/Shield-*-KD-*-Qwen3Guard-Gen-8B-FT-*) continue to resolve. The underlying training is a full-parameter fine-tune, not PEFT.

Test-set results (DIA-GUARD holdout, 181,874 samples)

Metric	Value
Accuracy	0.9515
Precision	0.9573
Recall	0.9476
F1	0.9507
ASR (unsafe caught)	99.33%

Metric	Value
SAE Accuracy (36,050 samples)	0.9634
SAE F1	0.9629

Comparison to our earlier LoRA-CE attempt

Variant	Holdout Acc	Holdout F1
LoRA-CE (our earlier attempt, deprecated)	0.5432 ⚠️	0.3545
Full-FT (aliwister, this repo)	0.9515	0.9507

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("jsl5710/Shield-Qwen3Guard-Gen-8B-FT-PEFT-CE")
model = AutoModelForCausalLM.from_pretrained(
    "jsl5710/Shield-Qwen3Guard-Gen-8B-FT-PEFT-CE",
    torch_dtype="bfloat16",
    device_map="auto",
)

prompt = (
    "Classify the following text as safe or unsafe.\n\n"
    "Text: How do I bake a cake?\n\n"
    "Answer:"
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=8)
print(tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Model architecture

Base	Qwen/Qwen3Guard-Gen-8B
Params	8.37 B
Hidden	4096
Layers	36
Vocab	151,936

Citation

If you use this model, please cite the original aliwister upload in addition to the DIA-GUARD work:

@misc{aliwister_qwen3guard_8b_diaguard_2026,
  author = {aliwister},
  title  = {qwen3guard\_gen\_8b-diaguard},
  year   = {2026},
  url    = {https://huggingface.co/aliwister/qwen3guard_gen_8b-diaguard},
}