qwen3.6-27b-abliterated-journalist

Training

• Method: LoRA with Unsloth FastModel + TRL SFT, following the official Unsloth Qwen3.5 fine-tune recipe (canonical for Qwen3.6 too).
• Base model: huihui-ai/Huihui-Qwen3.6-27B-abliterated (multimodal, hybrid-thinking, dense 27B, qwen3_5 architecture)
• Dataset: tomvaillant/investigative-journalism-training (687 examples, OSINT methodology)
• LoRA config: r=16, alpha=16, dropout=0; targets q/k/v/o/gate/up/down projections; bias="none"; use_gradient_checkpointing="unsloth"
• Precision: bf16 (load_in_16bit=True; 4-bit QLoRA explicitly not recommended for Qwen3.5/3.6 per Unsloth docs)
• Optimizer: adamw_8bit, lr 2e-4, warmup 10 steps, 2 epochs, effective batch = 1 × 4 grad-accum
• Vision tower: frozen (finetune_vision_layers=False) — text-only LoRA, vision capability preserved byte-identical to base
• Final loss: 0.52 (step 300, epoch ~1.75)
• Task: investigative reporting assistance, OSINT methodology, verification, public-records research, source handling, and ethics

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

This model is intended for journalist-facing assistant workflows: investigation planning, OSINT tool selection, verification checklists, public-source research methods, and evidence-grounded drafting. Verify model outputs before use in reporting.

This was trained with Unsloth.