Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# llm-for-sec adapters
|
| 2 |
+
|
| 3 |
+
LoRA adapters for the security-aware LLM mission (decoder-only lineage).
|
| 4 |
+
|
| 5 |
+
- `qwen3-8b-sec-m5a-dpo-r7` (rank16, Base control arm) — W 0/4: ASR 80.36 / jb 75.00 / contract 0/168
|
| 6 |
+
- `qwen3-8b-sec-m5a-dpo-r8` (rank16, aligned bare) — X 0/4: ASR 78.57 / ORR 10.00 / jb 77.50 / contract 0/168
|
| 7 |
+
- `qwen3-8b-sec-m5a-dpo-r9` (rank64, three-lever: capacity+dose+adversarial-form pairs+amb half-dose, 600 steps)
|
| 8 |
+
|
| 9 |
+
Base model for r7: Qwen/Qwen3-8B-Base. Base for r8/r9: Qwen/Qwen3-8B (aligned).
|
| 10 |
+
Training configs in `configs/` of the source repo. Gate instruments frozen (sha-verified).
|