recordbrief-models

FieldScribe (RecordBrief tuned models)

Fine-tuned Qwen3.5 models that turn messy dictated speech transcripts into clean structured documents — fully offline, built for phones (llama.cpp GGUF, Q4_K_M).

What the tuning adds over stock Qwen3.5

• Verbatim fidelity: numbers, units, brand names, serials survive exactly ("118 psi", "Carrier 58STA090") — measured 10/10 on held-out sets.
• No invention: missing details become [not stated] instead of hallucinated recommendations (stock base invented in 7/30 reports).
• Spoken self-corrections: "three... no wait, four inches" and "correction to what I said earlier" → only the corrected value is kept.
• Noise filtering: small talk, emotional rants and fillers are dropped; the factual core is extracted.
• Two output modes: home-inspection report (InterNACHI sections, [Safety]/[Repair]/[Monitor] tags) and universal structuring (topic headings in the language of the dictation).
• 6 dictation languages: en, ru, de, pl, es, fr (+uk via transcription).
• Injection-resistant: instructions inside the transcript are treated as data, not commands.
• GBNF-native: trained to a grammar-constrained output shape for guaranteed structure with llama.cpp.

Trained with Unsloth (QLoRA r=64, NEFTune) on 6,000 synthetic dictation→document pairs; shipped inside the RecordBrief Android app.