Blame for README.md

---
license: other
task_categories:
- text-generation
tags:
- board-games
- reasoning
- self-distillation
- opsd
- mcts
pretty_name: SPSD-Variants OPSD (grounded MCTS teacher context, 45 variants)
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl.gz
  - split: heldout_move
    path: frozen_probes/move.jsonl.gz
  - split: heldout_state_qa
    path: frozen_probes/state_qa.jsonl.gz
---

# SPSD-Variants-opsd

Grounded on-policy self-distillation (OPSD) teacher-context dataset over **45
board-game rule variants** (5 families × 9: connect4, domineering,
simplified_first_attack, simplified_othello, tic_tac_chess), derived from
trained MuZero/EfficientZero checkpoints (plan-528 v2).

Each row is a decision-state task (a move choice or one of six auxiliary
state-QA tasks). The **`privileged_context`** is the teacher signal: **grounded
natural-language reasoning** that discovers the best move over the tempting
alternative and verifies it against the opponent's best reply, grounded in fresh
MuZero-net values that choose the wording but are **never printed**, and never
naming the search engine. The student-facing `prompt`, the `completion`, and the
`target` are unchanged from the underlying pools — only `privileged_context`
carries the grounded reasoning.

## Contents

Files are gzip-compressed JSON Lines (`datasets` loads `.jsonl.gz` transparently).

| file | rows | purpose |
|---|--:|---|
| `train.jsonl.gz` | 213,406 | training split (170,725 move + 42,681 state-QA) |
| `frozen_probes/move.jsonl.gz` | 6,840 | held-out move probe (disjoint from train) |
| `frozen_probes/state_qa.jsonl.gz` | 40,841 | held-out six-task state-QA probe |
| `manifest.json` | — | full provenance, per-variant counts, balance, census fingerprint |

Per-family subsets are obtained by filtering `metadata.game_name` /
`metadata.family` (the five families are connect4, domineering,
simplified_first_attack, simplified_othello, tic_tac_chess).

State-QA tasks (six): `occupancy`, `legality`, `threat_count`,
`legal_action_count`, `legal_action_enumeration`, `successor_state`.

## Row schema

- `prompt` — student-facing board task (chat messages).
- `completion` — `Final answer: \boxed{…}`.
- `privileged_context` — grounded teacher reasoning + the boxed answer (teacher-only).
- `target` / `accepted_answers` — verifier + canonical answer.
- `metadata` — `task`, `game_name`, `decision_state_hash`, checkpoint provenance
  (path/alias/SHA/env-step, algorithm, `num_simulations`, backend), and
  `replay_evidence_horizon` (`rich` vs a conservative-fallback mode).

## Provenance & method

- Sources: the trained `simplified5` othello-variant EfficientZero checkpoints;
  36 variants via `random-prefix-full-game` checkpoint self-play, 9 via
  exhaustive real-env enumeration.
- Grounding: one fresh **50-simulation** MCTS search per decision state; a
  separation gate (min root gap 0.05) refuses a distinction the root value does
  not support and falls back to conservative prose (no row is ever dropped).
- Move coverage: ~54% rich grounded traces, ~46% conservative fallback.
- Seed 0. Verified corpus-wide: **no** raw search stats or engine names leak into
  `privileged_context`; state-QA rationales are array-free.

## Caveats

- **Text only** (`tokenizer_hash=not_tokenized`): tokenize and build the
  OPSD-ready rows for your trainer.
- 6 connect4 variants (`exact_four`, `holey_four`, `no_center_opening`,
  `one_diagonal_disabled`, `orthogonal_only`, `outer_opening`) carry a declared
  `legality` shortfall (positive-skewed legality QA — illegal-handle states are
  rare there); the pooled answer-class balance stays ≤ 0.65.
- Global per-task counts are balanced per family, not equal across the corpus
  (reported, not enforced).

Private research dataset.