qwen3-4b-instruct-apparmor-log-to-rules-lora

qwen3-4b-instruct-apparmor-log-to-rules-lora

LoRA adapter for Qwen/Qwen3-4B-Instruct-2507 that reads AppArmor audit log denials and suggests the policy rules needed to allow the required access. Trained on real-world AppArmor profiles, covering both exact-path and generalized rules. Generated rules are a draft for human review -- verify them before applying to an enforced security policy, since overly permissive rules can weaken security. Apply with peft.PeftModel on the base model, or merge for standalone use.

Model Details

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Finetuning method: lora
• Library: llm-lab-kit (thin wrapper over transformers / peft / trl)

Training

• Dataset: local (/root/.cache/huggingface/hub/datasets--Vjeong--apparmor-audit-to-rules/snapshots/e9638a9019ebd2ca0fd530a4827c919ddcc271a0/data/train.jsonl)
• Precision: bfloat16
• Learning rate: 0.0002
• Epochs: 2.0
• Effective batch size: 16 (micro=2 x grad_accum=8)
• Scheduler: cosine (warmup_ratio=0.03)
• LoRA: r=16, alpha=32, dropout=0.05
• Source commit: 73a2945

Metrics

{
  "train_runtime": 3276.5012,
  "train_samples_per_second": 3.097,
  "train_steps_per_second": 0.194,
  "total_flos": 1.9053299486872166e+17,
  "train_loss": 0.1660708155740732,
  "epoch": 2.0,
  "acceptance": {
    "gates": {
      "G1_rule_f1": "PASS",
      "G2_coverage": "PASS",
      "G3_syntax": "PASS",
      "G4_no_regression": "PASS"
    },
    "overall": "PASS",
    "measurements": {
      "before": {
        "rule_f1_concrete": 0.3821428571428572,
        "rule_f1_all": 0.2600976430976431,
        "precision": 0.2768703703703704,
        "recall": 0.253,
        "coverage_recall": 0.4577777777777778,
        "syntax_rate": 0.74
      },
      "after": {
        "rule_f1_concrete": 0.8537280701754387,
        "rule_f1_all": 0.7661989121989122,
        "precision": 0.7663174603174603,
        "recall": 0.7662222222222222,
        "coverage_recall": 0.8968888888888888,
        "syntax_rate": 0.9933333333333333
      },
      "general_acc_before": 0.49666666666666665,
      "general_acc_after": 0.5,
      "general_acc_drop_pp": -0.3333333333333355,
      "trainable_pct": 0.2924071978366944
    }
  }
}

Usage

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "Vjeong/qwen3-4b-instruct-apparmor-log-to-rules-lora")
tokenizer = AutoTokenizer.from_pretrained("Vjeong/qwen3-4b-instruct-apparmor-log-to-rules-lora")

messages = [{"role": "user", "content": "Hello!"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
prompt_len = inputs["input_ids"].shape[-1]
print(tokenizer.decode(outputs[0][prompt_len:], skip_special_tokens=True))

License

apache-2.0