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

Gemma 4 E2B Social Spain v12 Q4KM GGUF

Public experimental GGUF export for local llama.cpp / LM Studio testing.

This is a merged and quantized checkpoint built from google/gemma-4-E2B-it plus the v12 Social Spain recovery LoRA, then quantized to Q4_K_M.

Status

This is not a production social-services model and must not be presented as legal, medical, or professional social-work advice.

The recommended architecture is:

local GGUF + dated RAG + privacy filter + deterministic validators + repair/abstention + human escalation

The public demo of that wrapper is:

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

About Gemma 4 MTP

Google released Gemma 4 Multi-Token Prediction drafters as an inference-speed optimization. The MTP path for this project is currently documented and tested as a Transformers backend using:

google/gemma-4-E2B-it + apol/gemma4-e2b-social-spain-v12-recovery-lora-public
assistant: google/gemma-4-E2B-it-assistant

Do not assume this GGUF has MTP acceleration in LM Studio or llama.cpp until the exact runtime supports Gemma 4 MTP drafters and is benchmarked. The GGUF remains the simple local runtime; MTP is an additional experimental Transformers path.

Files

• —gemma4-e2b-social-v12-Q4_K_M.gguf: quantized model.
• —conversion_provenance_sanitized.json: conversion provenance.
• —merge_lora_provenance_sanitized.json: merge provenance.
• —MANIFEST.json: file sizes and SHA-256 hashes.
• —NOTICE: Gemma redistribution notice.

Local Use

In LM Studio, download gemma4-e2b-social-v12-Q4_K_M.gguf and load it as a local GGUF model.

Example llama.cpp smoke:

llama-cli \
  -m gemma4-e2b-social-v12-Q4_K_M.gguf \
  -sys "Eres un asistente social. Responde sin razonamiento oculto." \
  -p "Responde en espanol claro. Que datos personales no debe compartir una persona migrante por chat?" \
  -n 96 --ctx-size 512 --threads 8 --batch-size 32 \
  --single-turn --reasoning off --no-warmup --simple-io --log-disable

Safety Use

For social-assistance scenarios:

• —do not ask users to share full passport/NIE numbers, exact address, document scans, or third-party documents in chat;
• —do not invent phone numbers, addresses, URLs, deadlines, or requirements;
• —use dated official/local sources for mutable facts;
• —escalate minors, violence, exploitation, homelessness tonight, medical urgency, and self-harm to qualified human support;
• —keep human-in-the-loop review for any high-risk case.

License And Use Restrictions

This artifact is distributed under the Gemma Terms of Use and is subject to the Gemma Prohibited Use Policy:

• —https://ai.google.dev/gemma/terms
• —https://ai.google.dev/gemma/prohibitedusepolicy

This model is a modified derivative of Gemma. Modified files and generated artifacts are provided with prominent notices in this model card and NOTICE.

Do not use this model for prohibited uses, including unauthorized legal, medical, social-welfare, or other professional practice, automated decisions affecting material rights or well-being, harmful content, or processing private personal information without required rights and consent.