Warrior-SFT

Supervised fine-tuning data for an agent that plays a seat in a turn-based game over the Warrior protocol: read a rendered position, use an RL policy's recommendation, choose one legal action, and say why.

5,423 train / 582 validation rows, one row per decision, split by match rather than by row.

The shape of a row

system     the Warrior protocol system prompt (game-agnostic; the rules arrive in the payload)
user       the rendered position -- rules briefing, ASCII board, private notes, the seat's own
           recent decisions with justifications, other players, capture points, table chat, and
           the legal action list with the RL policy's probability mass inline
assistant  one take_action tool call: {action_id, justification}
tools      the take_action schema, with action_id enumerated to exactly the legal set

The recommendation is rendered on the actions themselves, not fetched through a tool:

move_9_9: Move to (9,9)  [expert 93%; leaves you at least 1 tile from the point at (9,8), 2 closer]
move_8_9: Move to (8,9)  [expert 7%; leaves you at least 1 tile from the point at (9,8), 2 closer]
move_4_9: Move to (4,9)  [leaves you at least 3 tiles from the point at (6,6), 3 further]
atk_s1:   Attack Mas (seat 1) at (10,2)  [hit chance 54%; 5.66 tiles away, target has 6 health]
end:      End your turn

An earlier version delivered the same numbers through a consult_expert tool. That cost a second round trip per decision — 6s instead of 3s, and 14 minutes per match on a protocol that asks ~90 decisions per match — and only helped on the turns the model chose to ask. Inline is unconditional and free.

Where it comes from

engine	Raifu Wars, a turn-based strategy game, driven headless
policy	RaifuWars-RL-ActionScorer-SelfPlay, 57,730 parameters
collection	144 matches, 6 boards, 8,930 decisions, 0 retries and 0 errors
boards	Arboretum, Crossroads, Dustbowl, Glacier, Cornfield, Islands

The policy wins 69% / 50% / 31% across Crossroads, Arboretum and Islands, where a fair share for one seat of four is 25%.

Action mix

move        2,274        tier_choice   334
attack      1,056        rush          282
end_turn      966        reload        271

end_turn is capped at 15% of rows rather than left at its natural third. It is the least informative decision — ending a turn has no structured reason behind it, so most of those rows justified themselves with nothing but the recommendation's confidence. Capped rather than dropped, because knowing when a turn is over is a real skill and a model that never ends its turn stalls the match.

Two decisions worth knowing about before you train on this

The justifications are DERIVED, never generated. They are built from the same structured fields the game exposes — does this destination sit on a capture point, is this the best hit chance offered, does this tier-up end the match — and not from a language model asked to explain a move it did not make. A generated rationalisation is a plausible sentence with no causal link to the choice, and training on those teaches a model to produce confident reasoning that does not track what it is doing. The cost is that they read like templates, because they are: 824 distinct strings over 5,423 rows.

Prompts come from the live renderer. The generator imports the sidecar's own _render_state and _render_actions rather than reimplementing them, so a row cannot drift from what a model is served at inference. A model trained on a layout it is never served is a model trained on nothing.

What this teaches, and its ceiling

Every target is the RL policy's own choice. A model trained on this is a wrapper: it should approach the policy's strength and will not exceed it. That is the point of a first dataset and it is also its limit.

Knowing when the recommendation is WRONG cannot be learned here, because there is no example of a good override in it. That is measured rather than assumed: across 2,872 decisions, in all 22 positions where the rules prove what the right answer is (a tier-up that ends the match outright), the policy already picks it. A dataset of good overrides needs a source of good overrides, and one does not exist yet.

Limitations, stated plainly

• Board imbalance. Arboretum is 36% of the collected decisions; Islands and Cornfield are ~5% each, because those two arms collected 9 matches instead of 24.
• 91.2% of targets are the recommendation's top choice. The remainder are positions where the recorded action differs from the current argmax. Those rows teach "sometimes deviate" without teaching when.
• One engine. The protocol is game-agnostic and the system prompt carries no rules, but every row here is Raifu Wars. Whether this transfers to another game over the same protocol is untested.
• No chat rows. The protocol supports a chat action and the seat can talk; the RL policy cannot, so none of these rows exercise it.
• select_tile appears once. One example of an action type teaches nothing.

Reproducing

python tools/make_warrior_sft.py "data/sft-tr-*.jsonl" \
    --expert runs/ppo-selfplay/last.pt --out data/warrior-sft