mahendra1980

triage-qwen3-8b-public-v5

README

License: apache-2.0

Intended use

Triage first pass only. It fast-labels an alert so a human (or a frontier model) sees the shortlist. In AEGIS, a frontier model remains the judge — this adapter never makes the final call on its own.
Not a standalone IDS/WAF, not a detection engine, and not a substitute for authorization or logging.

Training

Table

Base	`Qwen/Qwen3-8B` (Apache-2.0)
Method	4-bit NF4 QLoRA + gradient checkpointing (unsloth)
LoRA	r=16, alpha=32 on q,k,v,o,gate,up,down
Trainable params	43,646,976 of 8.23B (0.53%)
Seq len	2048
Batch	2 × 8 grad-accum (effective 16)
Epochs / LR	2.0 / 2e-4
Data	42,174 train + 7,826 val (synthetic / public corpus)
Final train loss	0.234

Held-out gate (eval)

Table with columns: Suite, Score, Note
Suite	Score	Note
historical	33 / 37	4 misses (gate allows ≤ 5)
injection	9 / 9	0 prompt-injection leaks
agentic	10 / 13	tracked
endpoint	8 / 8	tracked
tls	12 / 13	tracked
Result	PASS

Known misses (honest): none drop a malicious case to benign. The silent misses are severity slips — suspicious → benign (EVAL-021, AGT-E07, AGT-E10) and under-rated malicious → suspicious (EVAL-008, AGT-E09, TLS-E04), plus a couple of over-calls. This is why it is a first pass with a frontier judge, not a final verdict.

Usage

python

from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", device_map="auto", load_in_4bit=True)tok  = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")model = PeftModel.from_pretrained(base, "mahendra1980/triage-qwen3-8b-public-v5")

In AEGIS:

bash

LLM_BACKEND=local AEGIS_LOCAL_4BIT=1 \AEGIS_MODEL=mahendra1980/triage-qwen3-8b-public-v5   # triage first pass; frontier model judges

License

Adapter released under Apache-2.0, matching the Qwen3-8B base. You must also comply with the base model's license.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

mahendra1980

Model Tree

Base

Qwen/Qwen3-8B

Adapter

this model

Input Modalities

Text

Output Modalities