apol/gemma4-social-spain-public-artifacts

Gemma 4 Social Spain Public Artifacts

Public, sanitized subset of the Gemma 4 Social Spain research artifacts.

This repository is intended to share the useful, low-risk parts of the work:

• —the public Gradio demo wrapper;
• —deterministic wrapper evaluations;
• —safety validators and repair/abstention scripts;
• —release gates and mobile/local-first deployment notes;
• —evaluation and product-upgrade reports.

What Is Not Included

This public subset intentionally excludes:

• —model weights and GGUF files;
• —LoRA adapters;
• —complete training JSONL files;
• —GPT-5.5 teacher outputs and training rows;
• —raw student candidates and failed outputs;
• —RAG chunks or source-derived excerpts;
• —review queues;
• —upload logs, local caches, bytecode, tokens, and raw source snapshots.

Those artifacts remain private pending licensing review, expert review, and release gating.

Current Demo

Public Space:

https://huggingface.co/spaces/apol/gemma4-social-spain-demo

The Space is a deterministic safety/runtime demo. It does not load private model weights. It demonstrates the intended production architecture:

privacy filter -> dated RAG -> local Gemma-derived model -> validators -> repair/abstain/escalate -> human handoff

The Space also includes an optional fine-tuned GGUF panel. It is off by default for stability, but can be enabled in a local clone with ENABLE_GGUF_INFERENCE=1 and llama-cpp-python.

It now also documents an experimental Gemma 4 MTP path. MTP pairs the v12 LoRA target with google/gemma-4-E2B-it-assistant for speculative decoding in Transformers. This is for latency benchmarking only; safety behavior still comes from the wrapper, validators, repair/abstention layer, and human handoff.

Current Recommended Runtime

The recommended internal runtime remains:

v12 GGUF + dated RAG + validators + repair/abstention + human escalation

Public model artifacts:

apol/gemma4-e2b-social-spain-v12-q4-k-m-gguf-public
apol/gemma4-e2b-social-spain-v12-recovery-lora-public

The v13/v14 experiment branches are private/unpromoted; v12 is the public recommended release because it is the strongest verified local-first path.

Safety Boundary

This is research material. It is not legal, medical, or professional social-work advice. High-risk use with refugees, migrants, minors, violence, exploitation, health urgency, housing, legal status, or social benefits requires qualified human review, dated official/local sources, privacy governance, and incident handling.

Key Files

• —demo_space/: public Gradio wrapper demo.
• —scripts/eval_demo_wrapper.py: deterministic wrapper eval.
• —scripts/validate_social_outputs.py: output validators.
• —scripts/repair_social_outputs.py: repair/abstention layer.
• —scripts/run_public_gguf_demo.py: small local fine-tuned GGUF smoke demo.
• —scripts/benchmark_mtp_runtime.py: target-only vs target+MTP benchmark.
• —scripts/verify_hf_public_release.py: public Hub release verifier.
• —docs/mtp_runtime.md: notes for the experimental Gemma 4 MTP runtime.
• —reports/demo_product_upgrade_2026-05-04.md: product-style demo upgrade.
• —reports/demo_wrapper_eval_2026-05-04.md: latest wrapper eval result.
• —reports/mtp_benchmark_attempt_2026-05-05.md: MTP dry-run and blocked micro-benchmark attempt.
• —reports/mtp_benchmark_cpu_small_2026-05-05.md: completed CPU target-only vs target+MTP benchmark.
• —docs/release_gates.md: release gates before production/public claims.
• —docs/mobile_litert_deployment.md: mobile deployment notes.