gemma-4-E4B-it-jurilix-ner-lora

Gemma 4 E4B · Jurilix French entity extraction

This is a GGUF LoRA adapter, not a standalone model. It is intended for the anonymization.ner entity-extraction route in Jurilix. Apply it at strength 0.75 for that route and explicitly send strength 0 for other routes.

Exact model and runtime

• Base: google/gemma-4-E4B-it-qat-q4_0-gguf, revision 4b4a2c1d584be7264f87aac328a1bc739ce81b6c, file gemma-4-E4B_q4_0-it.gguf.
• Base SHA-256: 676c35070db6dbe52f93e9c864ee0fba4eddea94b9c875d9cb10daff453fbaee.
• Adapter: gemma-4-E4B-it-jurilix-ner-lora-f16.gguf, 69,798,976 bytes.
• Adapter SHA-256: f8a9cacef5fd4e65a7de337b01d4077103b3eeecee2b55834866388fbec10239.
• Required tested runtime: b10956-jurilix1. The published source recipe pins upstream llama.cpp and a Gemma JSON parser fix.
• Tested context: 16,384 tokens; macOS arm64 Metal. These results do not qualify GPU performance or inference quality on other platforms.

Start llama-server with --lora ADAPTER.gguf --lora-init-without-apply --cache-ram 0. For entity extraction, include "lora": [{"id": 0, "scale": 0.75}] in each request. For every other route, include "lora": [{"id": 0, "scale": 0}].

Disabling the RAM prompt cache is required for this tested runtime: its saved KV states do not include adapter identity/strength, so reusing them across strength changes can corrupt a comparison or a subsequent response. The slot cache remains available and is invalidated when the adapter strength changes. Do not use the unpatched upstream b10956 build as an equivalent runtime.

Development comparison

Matched application-level captures used the same frozen prompts, weights, context, native runtime and cache policy. The baseline had no adapter loaded. The metric below measures annotated identifying mentions fully covered by the application's final replacements, with user approval simulated.

None of these panels introduced a newly incomplete annotated mention relative to the base. Hard-panel exact typed occurrence F1 increased from 0.8256 to 0.9760. At strength 1, a separate negative panel gained an incorrect quantity proposal; 0.75 avoids that new error while retaining the measured coverage improvements. The three baseline public-agency false positives remain on the negative panel (11 protected characters in both arms). A 32-request native switching probe restored identical base response text after both nonzero-to-zero transitions on each of eight previously captured development inputs.

These are synthetic development annotations, not an independently reviewed or unseen test set. Some panels may overlap prior development or training sources. Automatic scores do not establish generalization, perfect anonymization or legal accuracy. Existing mistakes remain, including public-agency acronyms incorrectly proposed for masking and some incomplete identifying mentions. Review each proposed replacement before exporting a document.

With the adapter loaded but explicitly set to zero, 47 native requests covering 34 non-NER feature cases and three long structured summaries reproduced the base response text exactly. This is retention evidence; existing base-model quality failures on those other routes remain. A subsequent request-exact replay also checks the extraction captures against the newer application pipeline without claiming additional native inference. That replay improves the hard-panel base to 296/356 (707 exposed characters), while the adapter remains at 351/356 (36 exposed), with no newly incomplete mentions.

The adapter weights are unchanged from the trained checkpoint. The adjacent .provenance.json records training-source revisions, conversion metadata and the byte-preserving release rename. Private training examples are not included.