qwen3.6-27b-abliterated-journalist-merged

Files

• 15 safetensors shards (fp16, ~55GB total)
• tokenizer files
• chat template (chat_template.jinja)
• vision projector weights (byte-identical to base — vision tower was frozen during training)

Training

• Adapter: tomvaillant/qwen3.6-27b-abliterated-journalist
• Method: LoRA (Unsloth FastModel + TRL SFT) merged into bf16 via save_pretrained_merged. Recipe follows the official Unsloth Qwen3.5 fine-tune guide: bf16 LoRA, r=16/alpha=16/dropout=0, use_gradient_checkpointing="unsloth", optim="adamw_8bit".
• Dataset: tomvaillant/investigative-journalism-training (687 examples, OSINT methodology)
• Vision tower: frozen during training; merged weights byte-identical to upstream Huihui base

Sources And Attribution

Training data: tomvaillant/investigative-journalism-training — 687 instruction/response pairs synthesized by Claude Opus 4.6 (Anthropic) from the Buried Signals OSINT and investigative-journalism corpus: OSINT Navigator tool data, Indicator Media briefings, Buried Signals investigative skills, GIJN, Bellingcat, Verification Handbook 3, SPJ Code of Ethics, RCFP, and public manuals from UNESCO, Al Jazeera Media Institute, CiFAR, CIPE, and EJF/TEMPO Institute.

See the dataset card for the full source list, licenses, and per-partner attribution.

Intended Use

Direct loading via Transformers or vLLM for journalism and OSINT workflows. For desktop/laptop inference, prefer the GGUF variant. Treat outputs as leads, not verified findings.