Field	Value
Method	LORA
Precision	fp16
Quantization	LoRA + 4-bit (bnb nf4)
Mode	steps
Num Epochs	1
Max Steps	3500
Eval Steps	500
Save Steps	500
LR	0.0001
Max Length	768
per_device_batch_size	1
grad_accum	8

Evaluation (greedy, fixed-length decode)

Metric	Score
BLEU-4	1.46
ROUGE-L	13.40
F1 (token-level)	25.38
chrF++	20.26
BERTScore F1	82.65
Perplexity	16.85

Notes: We normalize whitespace/punctuations, compute token-level P/R/F1, and use evaluate's sacrebleu/rouge/chrf/bertscore.

How to use

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
tok  = AutoTokenizer.from_pretrained("nhonhoccode/qwen3-0-6b-cybersecqa-lora-4bit-20251112-1252")
base = AutoModelForCausalLM.from_pretrained("unsloth/Qwen3-0.6B")
mdl  = PeftModel.from_pretrained(base, "nhonhoccode/qwen3-0-6b-cybersecqa-lora-4bit-20251112-1252")  # Loads LoRA adapter
prompt = tok.apply_chat_template(
    [{"role":"system","content":"You are a helpful assistant specialized in cybersecurity Q&A."},
     {"role":"user","content":"Explain SQL injection in one paragraph."}],
    tokenize=False, add_generation_prompt=True
)
ids = tok(prompt, return_tensors="pt").input_ids
out = mdl.generate(ids, max_new_tokens=128, do_sample=False)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

Intended Use & Limitations

Domain: cybersecurity Q&A; not guaranteed to be accurate for legal/medical purposes.
The model can hallucinate or produce outdated guidance—verify before applying in production.
Safety: No explicit content filtering. Add guardrails (moderation, retrieval augmentation) for deployment.

Reproducibility (env)

transformers>=4.43,<5, accelerate>=0.33,<0.34, peft>=0.11,<0.13, datasets>=2.18,<3, evaluate>=0.4,<0.5, rouge-score, sacrebleu, huggingface_hub>=0.23,<0.26, bitsandbytes
GPU: T4-class; LoRA recommended for low VRAM.

Changelog

2025-11-12 12:53 — Initial release (LORA-4bit)