Sandy-sys/osint-argus-gemma4-qat-w4a16

OSINT Argus — Gemma 4 12B QAT, unified news + social OSINT (W4A16, tri-modal)

All numbers on this card are measured on the deployed int4 artifact (previous cards quoted bf16-checkpoint numbers; the int4 reality is ~10pp lower on schema tasks — that discrepancy was the programme's audit finding A1, now fixed).

Unified build (2026-09): full-parameter QAT on the unified trunk — no adapter, no merge. Lineage: social-media soup (25-Aug) -> news track v20/v21 (120,747-record corpus: 40% news core — article veracity in Bengali/Hindi/Urdu/25-language x-fact + Reddit-format Fakeddit, FEVER-family claim adjudication, XL-Sum summaries in 9 languages, article political bias — plus full social/multimodal replay). Exported through the corrected int4 grid (earlier revisions shipped fq(fq(W)); fixed 31-Aug, verified lossless: baked-checkpoint == export). Selected over v20/v22 and the deployed soup-int4 on combined social/multimodal + news gates.

Text gates — 200 items/task, bare decoding, deployed int4

metric	stock Gemma-4 QAT (bf16)	this model (int4)
schema_lock (11-key OSINT record)	0.0	0.875
NER micro-F1	0.6163	0.7
sentiment acc	0.76	0.775
claim refuted-F1	0.8125	0.88
halluc-detect F1	0.3981	0.4786
grounded-answer schema	0.9536	1.0
general MCQ %	59.0	52.5

NLI-judged faithfulness (mDeBERTa-v3 multilingual, P>=0.70, all-sentences-entailed): 0.059602649006622516 (strict sentence-level entailment of abstractive answers — reported for transparency; schema-validity above measures format compliance only)

News gates — held out by content-hash firewall, never trained on (new in this build)

metric	this model
news veracity — misinfo accuracy	0.2335
news claim — macro F1 / acc	0.3956 / 0.4643
news bias — accuracy	0.3667
news summary — schema / filled	0.425 / 0.475

Multimodal gates — int4

metric	this model
vision OCR — CER ↓	0.6368
vision misinformation — accuracy	0.525 (majority baseline 0.5083)
vision hostility — accuracy	0.4434 (majority baseline 0.6321)
audio ASR — CER ↓	0.6835
audio language-ID — accuracy	0.1833 (majority baseline 0.175)
audio abusive-speech — accuracy	0.4 (majority baseline 0.5417)

Serve

python -m sglang.launch_server --model-path <this-repo> \
  --quantization compressed-tensors --mem-fraction-static 0.45 --context-length 8192

Training code, gate reports, gap audit, hold/ship decisions: shubro18202758/osint-argus on GitHub.

Card generated 2026-09-03 from this artifact's own int4 gate reports.