mhndayesh/gemma-4-12B-netsec-expert-GGUF

⚙️ Recommended runtime settings — gemma-native sampling temperature 1.0, top_k 64, top_p 0.95, min_p 0.01 (the `min_p 0.01` floor prevents the reasoning-loop empty-answer issue), context length ≥ 16k (32k recommended), and a generous max_tokens when running with thinking on. The Gemma-4 thinking path needs --jinja.

gemma-4-12B-netsec-expert (GGUF)

*Base `gemma-4-12B-it-QAT` with a Security & Networking FactBank baked into its chat-template. The model answers correctly about post-cutoff / breaking-change APIs in 7 security & networking libraries — not by fine-tuning, but by carrying a searchable bank of landmine facts that fires inside llama.cpp at inference time. Weights are untouched* (only the GGUF chat-template was rewritten); no external RAG service.

The model supplies the reasoning; the bank supplies the knowledge it was never trained on.

🔗 Full project — all experts, methodology, per-question transcripts, and benchmarks: [github.com/mhndayesh/experts-models](https://github.com/mhndayesh/experts-models)

What it fixes

114 curated landmine facts (post-cutoff · reverses-a-trained-habit · silent-failure) across: cryptography · OpenSSL 3 · paramiko 3 · urllib3 2 · volatility3 · yara-x · eBPF (BCC→libbpf). Examples the base gets wrong and this model gets right: RSA_new→EVP_PKEY_new, BN_is_prime_ex→BN_check_prime, FIPS_mode→EVP_default_properties_is_fips_enabled, SSL_CTX_new→SSL_CTX_new_ex, TripleDES→hazmat.decrepit, the volatility3 v3 plugin (PluginInterface+TreeGrid+run), yara-x base64/wildcard rule changes.

Libraries in the bank (7) — 114 facts total

Where the facts come from (mined sources)

Each library's facts were extracted from its migration guide / changelog (source targeting is the whole game — a migration guide, not release-note noise), then quote-verified against the source line:

• —cryptography changelog
• —OpenSSL 3 migration guide
• —eBPF BCC→libbpf migration guide
• —paramiko changelog
• —urllib3 v2 migration guide
• —volatility3 migration guide
• —yara-x differences doc

Full provenance (the mined source docs themselves) lives in the repo under `v2/extractor/experts/security-networking/sources/`.

Results (this model — hand-verified)

Same 48 landmine questions, base vs. this baked model, identical prompts (the bank injects in-engine):

Part of a 2B/12B/26B curve (e2b 39.6→66.7% · 12B 56.2→81.2% · 26B 77.1→93.8%): raw lift shrinks with size, error-closure rises. Full methodology, per-question transcripts, and caveats: github.com/mhndayesh/experts-models → v2/extractor/experts/security-networking/.

How to run

The bank lives in the chat-template, so retrieval needs the template applied — run on llama.cpp:

llama-server -m gemma-4-12B-netsec-expert-Q4_0.gguf --jinja --port 8080 --ctx-size 8192

Then query normally (the same prompt you'd send the base model — the bank fires automatically for covered topics). Sampling — use Gemma-native, not a bare low temperature: temperature 1.0, top_k 64, top_p 0.95, min_p 0.01 (the min_p floor prevents reasoning-loop empty answers).

Best accuracy (authority + thinking): send chat_template_kwargs={"enable_thinking": true} and a system prompt that tells the model the looked-up facts are verified and supersede its training. A reasoning model otherwise tends to "correct" an injected fact back to its trained prior — authority framing holds the fact. (The 118 KB template is under the LM Studio raw-load size wall, so LM Studio also works — but it ignores chat_template_kwargs, so use llama-server for the thinking-on mode.)

Limitations

• —Scoped to the 7 covered libraries. Outside them it's the base model.
• —Supplies knowledge, not reasoning. Multi-step transforms (e.g. some eBPF CO-RE conversions) can still fail even with the right fact retrieved.
• —Retrieval gate is token-based, with aliases. This bake includes the gate-alias fix — a natural or old name (e.g. "Volatility 3", RSA_new) also opens the tab; a wholly unrelated phrasing may still miss.
• —Numbers are hand-scored landmine tests, not a general coding benchmark.

Also in this project — GitChameleon 2.0 vs. the frontier

The same FactBank approach on a different, code-execution benchmark: a local 12B + bank next to the published GitChameleon 2.0 leaderboard.

⚠️ Not apples-to-apples — the container harness was NOT run. Frontier = the official 328-problem run in pinned Docker containers (arXiv 2507.12367). FactBank rows = a local, non-Docker run over the 249 problems that built (hand-verified, some 3.7→3.9 remapped, no RAG). Base-vs-baked is internally fair; the frontier column is a different measurement — "what neighborhood," not a ranking. Details: LEADERBOARD-COMPARISON.md.

Papers

The write-ups behind this project (PDFs, rendered on GitHub): Research Report · Idea · Technical · Verdict · Evidence Ledger

Provenance & license

• —Base: lmstudio-community/gemma-4-12B-it-QAT-GGUF (Q40). This model = that GGUF with `tokenizer.chattemplate rewritten to embed an inverted-index retriever + the bank (factbank.version 0.4.0`).
• —License: Google Gemma Terms of Use (license: gemma) — this is a gemma-4 derivative. The fact bank is from the FactBank project (see the repo LICENSE); mined sources keep their own licenses.
• —Source & method: github.com/mhndayesh/experts-models.