bdougie/pokemon-red-telemetry-sft

Pokemon Red Telemetry SFT corpus (sft_v3)

590 chat-format SFT examples generated deterministically (seed 42) from ~38k pokemon.game.v1 telemetry events emitted by an autonomous Pokemon Red agent (pokemon-kafka) and its training loop (empirical-evidence). Labels for the battle domains come from game RAM, not annotation.

This corpus was used to fine-tune a multi-task SmolLM3-3B LoRA in the empirical-evidence loop, which beat the base model on held-out game-RAM labels (win-prediction 1.00 vs 0.00, type-effectiveness 0.56 vs 0.00); that multi-task checkpoint is not yet published. (The separate bdougie/smollm3-forest-lora adapter was trained on a different, forest-crossing corpus — not this one.)

Configs / splits

config	contents
default	corpus.jsonl — all 590 examples, each row keeps its domain field (browse by domain here)
split	the modeling split: train (533, domain stripped) + validation (57, stratified 10%)

from datasets import load_dataset
full  = load_dataset("bdougie/pokemon-red-telemetry-sft")                 # default: corpus
split = load_dataset("bdougie/pokemon-red-telemetry-sft", "split")        # train + validation

Domains

domain	examples	source	label
narrator	236	milestone / map_change / discovery / battle_end events	template-generated one-sentence play-by-play (synthetic style labels)
battle-action	144	per-turn battle events of battles that were won	the action JSON the agent actually took (rejection-sampled on outcome)
move-choice	97	move_result rows	damage bucket vs enemy max HP; best-move picks where ≥2 moves observed per matchup
battle-outcome	72	battle_outcome rows	recorded win/loss → fight/flee recommendation
genome	41	rollout genome+fitness artifacts	above-median-fitness survival-parameter JSON per scenario

Files

• corpus.jsonl — all examples, each {"messages": [...], "domain": ...}
• train.jsonl (533) — domain stripped, feeds TRL SFTTrainer directly
• valid.jsonl (57) — stratified 10% split, keeps domain for per-domain eval
• stats.json — per-domain counts, skipped-line count, seed, corpus SHA-256

Generation

uv run python -m autotune.convert_telemetry \
  --pk-data ../pokemon-kafka/data --ee-data data/telemetry \
  --out data/sft_v3 --seed 42

Deterministic: same inputs + same seed → byte-identical corpus (corpus_sha256 in stats.json; converter source: autotune/convert_telemetry.py).

Known limitations

• Early-game Pokemon Red only (Pallet Town → Pewter City area); narrow species/type coverage.
• ~55% of battle-action prompts contain raw #NN species codes where the RAM reader had no resolved name (e.g. #71); the model sees these as opaque identifiers.
• Narrator labels are template-generated (5 seeded phrasings per event type) — stylistic, not human commentary.
• Heavy dedup: the agent grinds repetitive early-game states, so ~38k raw events collapse to 590 unique examples.