swohamkayastha/legacyscribe-9b-v3-lora

LegacyScribe 9B v3

Fine-tuned Qwen3.5-9B for conversational memory preservation — transforms spoken memories into formatted life story chapters. Supports English, Nepali, and mixed input.

Five-agent pipeline

• Questioner — warm follow-up questions to draw out more story
• Arc classifier — tags narrative stage (setup / tension / turn / meaning)
• Extractors — pulls person, place, time, emotion as structured JSON
• Reconciler — detects contradictions across memory notes
• Publisher — synthesizes notes into a warm first-person narrative passage

Usage (GGUF / Ollama)

ollama run hf.co/your-username/legacyscribe-9b-v3-gguf

Training

• Base: Qwen3.5-9B (4-bit, Unsloth + LoRA)
• LoRA r=16, alpha=32, 5 epochs
• Dataset: custom Nepali/South Asian memory examples """

with open("./models/legacyscribe-9b-v3-gguf/README.md", "w") as f: f.write(readme)