sharky-0.5B-v3 β€” archived phase9 candidate (REJECTED per pre-registered gate)

Mixed-corpus retrain (UNSW-NB15 + DAPT-2020 + CIC-IDS2017 60.5k + CTU-13 10k states, from base, 18,208 steps / 291,308 records, lr 5e-5). Not promoted β€” the production model remains luispoveda93/sharky-0.5B.

Results (phase9 5-gate battery β€” 3/5)

Suite v2 production v3 (this)
CTU-13 Botnet-42 (real malware) recall 7.4% 99.0% βœ“
Unseen 2018 benign recall (false-flag) 58% (34% FF) 97.2% (2.8%) βœ“
Locked (in-distribution) 0.9865 0.9808 βœ“
CIC-IDS2017 macro-F1 0.173 0.706 (target 0.85) βœ—
OOD ECE 0.33–0.51 S1 0.19 βœ—, S2b/S4 ≀0.015 βœ“

Why rejected: benign recall on CIC-IDS2017 is 0.36 β€” diagnosed as a benign-web shortcut (phase9/diag_s1_benign.json): web-destination benign flows are false-flagged 85%.

Full record: phase9/REJECTED.md Β· gate_v3.json Β· CONTRIBUTING.md

Serve exactly like the production model (POST /v1/systemone via the same shim); the same prefill-only battery applies. Useful as a real-malware detector today; not production-safe on benign web traffic.

Downloads last month
41
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for luispoveda93/sharky-0.5B-v3

Adapter
(4)
this model