Persona fine-tuning — your instance’s voice in the weights

Furoshiki already shapes personality at runtime — identity shell, emotions, needs, memory retrieval, trait observations, and situational roles all land in the prompt every turn. Persona FT is the complementary track: a local LoRA style adapter trained on real exchanges so the base model’s default prose habits bend toward how this Furoshiki actually talks on Telegram. Facts, user profile, mood, and RAG stay out of the weights; they continue to flow through SQLite, Chroma, and prompt assembly.

Goal

Build a per-instance conversational style model that operators can refresh as new gold replies accumulate. The system should:

• Ground training targets in real assistant messages (or operator-edited gold), not free-text self-observations alone.
• Keep a human-in-the-loop curation queue so bad or off-brand turns never silently enter a train set.
• Retrain from the base instruct checkpoint every cycle — never stack adapter-on-adapter, so drift stays bounded.
• Gate publish with held-out test loss vs the current champion adapter before Ollama registration.
• Plug into the existing stack as one model tier (e.g. ollama::furoshiki-persona:latest) while the rest of Furoshiki stays unchanged.

Where it fits in Furoshiki

Over months, the prompt layer tells the model what matters today; the LoRA layer makes the same instructions sound more like the relationship you have already built — shorter clauses, warmth, directness, humor timing — without duplicating Chroma into weights.

End-to-end process

CAPTURE  Telegram /persona keep · /persona capture on
         │   Post-conversation LLM mining (source=automatic, quality-gated)
         │   Linked exports from personality tool logs + curated session logs
         ▼
CURATE   Dashboard Persona FT or CLI pending/promote/reject/edit
         │   Rows: situation + user line + assistant gold (+ trait / style note)
         ▼
EXPORT   furoshiki persona-ft export → cycles/<id>/data/{train,valid,test}.jsonl
         │   Chat template: fixed system prompt + structured user brief + assistant reply
         ▼
TRAIN    mlx_lm LoRA on base (default: Llama 3.2 3B Instruct 4-bit)
         │   Small-data hyperparams (low LR, ~200 iters, rank 8)
         ▼
EVAL     Test loss vs champion; optional commitment quotes as human gold
         ▼
PUBLISH  Fuse adapter → Ollama tag (default furoshiki-persona)
         │   Promoted rows → status used for this cycle
         ▼
RUN      Point a model tier at ollama::furoshiki-persona:latest
         Full prompt pipeline unchanged — adapter only affects token sampling.

Capture sources

What each training example looks like

Examples are mlx-lm chat JSONL triples. The system message is a fixed persona training prompt ($FUROSHIKI_DATA/persona_ft/system_prompt.txt, must match Ollama SYSTEM after publish). The user message is a structured brief — situation, trait focus, optional style note (explicitly “do not recite”), sometimes a behavioral commitment — ending with “Respond in character — one short message.” The assistant target is the real or curated Telegram reply (clipped to a sane length), never the observation prose itself.

Curation and lifecycle (persona_ft_examples)

Every accumulating row has a named downstream — no dead-end table:

Dashboard Persona FT and repair API routes mirror the CLI: bulk promote/reject, inline edit of situation and messages, manual insert for gold pairs. This is the quality gate that keeps small-data LoRA from memorizing one bad evening.

Train, eval, publish

Artifacts

All under $FUROSHIKI_DATA/persona_ft/:

• config.json — base model, LoRA knobs, Ollama tag, eval ratio cap
• cycles/<id>/data/ — train / valid / test JSONL + manifest
• cycles/<id>/adapters/ — LoRA weights for the cycle
• champion/ — last adapter that passed eval (metadata in meta.json)

Defaults (operator-tunable)

• Base: mlx-community/Llama-3.2-3B-Instruct-4bit
• LoRA rank 8, alpha 16, learning rate 5e-6, ~200 iterations (anti-overfit for small sets)
• Publish only if test loss ≤ champion loss × max_test_loss_ratio (default 1.02), unless forced

CLI

Prerequisites: Apple Silicon Mac with sufficient unified memory (32 GB comfortable for 3B QLoRA), pip install -r requirements-persona-ft.txt, optional Ollama for publish. Canonical operator doc in the repo: docs/PERSONA-FT.md.

How this improves personality over time

1. Weeks 1–2 — prompt-only. Trait observations, roles, and memory do the heavy lifting; you capture standout Telegram turns with /persona keep.
2. First cycle — baseline adapter. After ~dozens of promoted pairs, export + train establishes a champion that tracks your curated voice.
3. Ongoing — dual feedback. Runtime systems record what felt authentic (observations, commitments honored); Persona FT records how it was said (assistant strings). Mining adds candidates; you still promote manually.
4. Periodic refresh. Each cycle retrains from base + full promoted corpus (plus export joins), compares eval loss, and swaps Ollama only when the candidate wins — so regression is detectable.
5. Composability. When you enable the persona tier for live chat, the model still receives the same behavioral cues, memory, and corrections; users perceive tighter alignment because phrasing matches history, not because facts were baked into weights.

Expected outcomes

• Success: Replies feel more like your established Telegram voice with less “generic assistant” cadence, while factual questions still defer to retrieval and tools.
• Not a goal: Replacing OpenRouter/cloud models entirely, storing private facts in the checkpoint, or automating promotion without operator review.
• Failure modes we design against: overfitting tiny sets (low LR, short iters, eval gate), training on observation text (export filters), adapter stacking (full retrain rule), silent quality decay (champion comparison).

Related systems

• Emotion state & needs — still injected every turn; not trained into LoRA.
• System architecture — prompt bands, memory, reflection loops.
• User guide — Persona FT — Telegram commands, dashboard, env toggles for auto-mine.
• Personality evolution (traits, auto-learned roles) — parallel track; see dashboard Personality evolution and docs/PERSONALITY-EVOLUTION-PLAN.md.