mhndayesh/gemma-4-12B-dataplane-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-dataplane-expert (GGUF)

*Base `gemma-4-12B-it` with an eBPF & the Networking Data Plane FactBank baked into its chat-template. The model answers correctly about post-cutoff / breaking-change APIs in 7 networking data-plane libraries — not by fine-tuning, but by carrying a searchable bank of 318 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 this expert is about

The Linux networking data plane — eBPF/XDP in the kernel, DPDK/VPP in userspace, and the controllers (Cilium, FRR) that program them — churns hard: the whole BCC→libbpf (CO-RE) shift, libxdp splitting out of the kernel tree, Cilium config keys renamed (tunnel→routingMode). This bank carries those breaking-change facts so the model stops answering from the old world.

Everything here is a landmine fact: post-cutoff, reverses-a-trained-habit, or a silent failure the model wouldn't otherwise catch. Every fact is quote-grounded to a real migration guide / changelog line. Examples the base gets wrong and this model gets right: the BCC→libbpf CO-RE conversion, libxdp's split from the kernel, Cilium's tunnel→routingMode, DPDK removed APIs, FRR operator changes.

Libraries in the bank (7) — 318 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:

• —eBPF BCC→libbpf migration guide
• —libbpf v1 notes
• —xdp/libxdp split doc
• —cilium upgrade guide
• —DPDK deprecation notices
• —VPP release notes
• —FRR changelog

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

Results (this model — hand-verified)

Same 47 adversarial landmine questions, base vs. this baked model, identical prompts (the bank injects in-engine). Config: thinking-on + authority framing, Gemma-native sampling.

\ of the answers the base got wrong, the fraction the bank fixed. Every answer was hand-scored* — an automated substring check miscounts in both directions (it fails a correct answer that names the old API as a contrast, and passes a semantically wrong one).

Honest note (12B): ~3 of 47 cases regressed — the base was already right and the injected fact + reasoning made it worse (the reasoning-paradox tax). Reported, not hidden.

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-dataplane-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) — this is how the numbers above were measured. Send chat_template_kwargs={"enable_thinking": true} and a system prompt telling 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.

Limitations

• —Scoped to the 7 covered libraries. Outside them it's the base model.
• —Supplies knowledge, not reasoning. A multi-step transform can still fail even with the right fact retrieved.
• —Retrieval gate is token-based, with aliases. This bake includes the gate-alias fix — a rename's OLD name (e.g. "CrackMapExec") also opens the tab — but a wholly unrelated phrasing may still miss.
• —Numbers are hand-scored landmine tests, not a general coding benchmark.

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`, gate-alias fix applied).
• —License: Google Gemma Terms of Use (license: gemma) — 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.