arubinochan-bot

Misskey のタイムラインを読み、ローカル LLM で「あるびのちゃん」として投稿する bot です。

To install dependencies:

bun install

To run:

bun run index.ts

Dry run without posting:

bun run index.ts --test

Useful runtime options:

bun run index.ts --test --harness speaker --politeness auto --sampling creative

• --harness speaker: default. Starts generation with あるびのちゃん:\n and parses the following body.
• --harness tagged: wraps output in a lightweight <post> envelope.
• --harness json: legacy JSON grammar harness.
• --politeness auto: default. Rephrases once only when the generated text does not look polite enough.
• --politeness always: always runs the style converter.
• --politeness never: disables style conversion.
• --sampling creative: default. Keeps the more idiosyncratic tone used by the bot.
• --sampling model-default: uses model-card-like sampling parameters for supported models.

Environment variables:

MISSKEY_ORIGIN=https://misskey.example.net
MISSKEY_CREDENTIAL=XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
MODEL="LiquidAI/LFM2-2.6B-GGUF:Q5_K_M"
HARNESS="speaker"
POLITENESS="auto"
SAMPLING="creative"

Evaluation without posting:

bun run evaluate.ts --all-env --candidates --runs 1 --output .evaluation-results.jsonl

Resume a crashed/interrupted evaluation:

bun run evaluate.ts --all-env --candidates --runs 1 --output .evaluation-results.jsonl --resume

See EVALUATION.md for the harness/model comparison and the current recommendation.