gemma4-e4b-abliterated-journalist

LoRA adapter for investigative journalism and OSINT workflows, fine-tuned from huihui-ai/Huihui-gemma-4-E4B-it-abliterated.

This is the adapter checkpoint. For local inference, use the merged or GGUF variants:

• gemma4-e4b-abliterated-journalist-merged
• gemma4-e4b-abliterated-journalist-GGUF

Training

• Method: QLoRA with Unsloth + TRL SFT
• Base model: huihui-ai/Huihui-gemma-4-E4B-it-abliterated
• Dataset: tomvaillant/investigative-journalism-training
• Task: assistant responses for investigative reporting, OSINT tool selection, evidence handling, verification, media law, and ethics

Sources And Attribution

Training examples were synthesized and curated from Buried Signals investigative journalism materials and public/open journalism references. Key source families:

• OSINT Navigator Tool Database: tomvaillant/osint-tool-database, a database of 7,524 OSINT tools assembled from open-source toolkit indexes and maintained by Tom Vaillant / Buried Signals.
• Indicator Media briefing tools: 101 OSINT tool entries and use-case descriptions from Indicator Media briefings, used with permission.
• Buried Signals skill repositories: OSINT methodology, Follow the Money, social media intelligence, OPSEC, evidence grounding, Spotlight investigator, and Spotlight fact-checker materials.
• Investigative journalism references: GIJN resources, Bellingcat guides, Verification Handbook 3, SPJ Code of Ethics, RCFP resources, and manuals from UNESCO, Al Jazeera Media Institute, CiFAR, CIPE, and EJF/TEMPO Institute.
• Synthetic instruction data: 700 instruction/response examples generated with Claude Opus 4.6 and grounded in the above corpus, with prompts emphasizing SIFT verification, source protection, proportionality, and evidence standards.

Full source register is maintained in the training repository as SOURCES.md.

Intended Use

This model is intended to support journalists and researchers with investigative planning, OSINT methodology, source triangulation, tool selection, and structured verification workflows. Outputs should be treated as leads and drafts, not as independently verified facts.

This was trained with Unsloth.