README.md

---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - chess
  - sft
  - instruction-tuning
  - reasoning
  - chess960
pretty_name: Chess SFT Training Data
configs:
  - config_name: default
    data_files:
      - split: train
        path: "tier*/*.jsonl"
  - config_name: tier0_tier0
    data_files:
      - split: train
        path: "tier0/*.jsonl"
  - config_name: tier1_perception
    data_files:
      - split: train
        path: "tier1/*.jsonl"
  - config_name: tier2_rules
    data_files:
      - split: train
        path: "tier2/*.jsonl"
  - config_name: tier3_tactics
    data_files:
      - split: train
        path: "tier3/*.jsonl"
  - config_name: tier4_evaluation
    data_files:
      - split: train
        path: "tier4/*.jsonl"
  - config_name: tier5_openings
    data_files:
      - split: train
        path: "tier5/*.jsonl"
  - config_name: tier6_endgames
    data_files:
      - split: train
        path: "tier6/*.jsonl"
  - config_name: tier7_planning
    data_files:
      - split: train
        path: "tier7/*.jsonl"
---

# Chess SFT Training Data

A supervised fine-tuning dataset for teaching language models to reason about
chess. It covers board perception, legal move generation, tactics, evaluation,
openings, endgames, and planning.

| | |
|---|---|
| **Total examples** | 15,100,000 |
| **Total size** | 18902.7 MB |
| **Format** | JSONL chat rows with `messages` |
| **Eval companion** | [Chess-Nut-Engine/chess-sft-eval](https://huggingface.co/datasets/Chess-Nut-Engine/chess-sft-eval) |

## Tier Overview

| Tier | Category | Tasks | Examples | Size |
|------|----------|-------|----------|------|
| 0 | Unknown | 1 | 300,000 | 556.2 MB |
| 1 | Perception | 19 | 6,240,000 | 8029.8 MB |
| 2 | Rules | 14 | 3,280,000 | 4298.8 MB |
| 3 | Tactics | 9 | 1,800,000 | 1759.3 MB |
| 4 | Evaluation | 3 | 600,000 | 447.9 MB |
| 5 | Openings | 3 | 120,000 | 101.1 MB |
| 6 | Endgames | 4 | 560,000 | 386.8 MB |
| 7 | Planning | 14 | 2,200,000 | 3322.7 MB |


## Loading

```python
from datasets import load_dataset

ds = load_dataset("Chess-Nut-Engine/chess-sft-corpus-4x", streaming=True)
ds_tier1 = load_dataset("Chess-Nut-Engine/chess-sft-corpus-4x", "tier1_perception")
```

## Row Schema

Every row is one JSON object:

| Field | Type | Meaning |
|---|---|---|
| `task` | str | Curriculum task id (e.g. `2.1_legal_move_gen`) — filter/join key |
| `tier` | int | Curriculum tier 1-7 |
| `fen` | str | Anchor position (FEN) |
| `is_chess960` | bool | Chess960 row |
| `messages` | list | Chat turns (`system`/`user`/`assistant`); `7.12_game_episode` rows are multi-turn |
| `metadata` | object | Task-specific provenance (expected answers, source ids, dedup identity) |

Filtering by task across configs:

```python
legal = ds["train"].filter(lambda row: row["task"] == "2.1_legal_move_gen")
```

## Data Sources

Rows are generated from Lichess games, Lichess puzzles, Lichess openings,
Lichess position evaluations, MATE rows, Syzygy tablebases, Polyglot opening
books, and generated Chess960 positions. Eval and benchmark FENs are excluded
from training with a blocklist (game-scoped: sibling positions of eval games
are blocked too).

## Provenance
- SDPO seed pack (2026-07-09): five new tier-7 tasks seeding RL-amplifiable reasoning behaviors, all engine/board-verified. 7.14_refute_and_switch (120k): tempting candidate -> its own engine-PV refutation with verified consequence -> 'Backtrack:' -> best. 7.15_composed_audit (120k): Check/Material/Captures/Hanging skill pipeline inline, then rated candidates. 7.16_prose_analysis (120k): grounded natural-language analysis in three styles (problem-focused / intuition-then-verify / comparison), phrase-pool diversity, zero unverified claims. 7.17_puzzle_episode (160k): full forced puzzle solutions as multi-turn episodes (opponent replies injected as user turns). 7.18_line_tracking_episode (120k): state the engine PV as a plan, execute it across turns ('On plan: <uci> was expected.'). Intended as an SFT top-up before GRPO/SDPO.
- Also refreshed via identity-dedup resume: 7.2_puzzle_solving, 7.11_history_best_move.

## Detailed File Listing

### Tier 0 - Unknown

| File | Task | Examples | Size |
|------|------|----------|------|
| `tier0/0.1_general_instruct.jsonl` |  | 300,000 | 556.2 MB |

### Tier 1 - Perception

| File | Task | Examples | Size |
|------|------|----------|------|
| `tier1/1.10_fen_row_application.jsonl` |  | 400,000 | 781.0 MB |
| `tier1/1.11_square_coordinates.jsonl` |  | 400,000 | 411.6 MB |
| `tier1/1.12_fen_rank_expansion.jsonl` |  | 400,000 | 397.4 MB |
| `tier1/1.13_fen_rank_cell_edit.jsonl` |  | 400,000 | 373.1 MB |
| `tier1/1.14_fen_board_edit.jsonl` |  | 400,000 | 523.7 MB |
| `tier1/1.15_material_inventory.jsonl` |  | 200,000 | 237.1 MB |
| `tier1/1.16_material_piece_counts.jsonl` |  | 200,000 | 216.0 MB |
| `tier1/1.17_material_value_totals.jsonl` |  | 200,000 | 224.7 MB |
| `tier1/1.18_material_balance_trace.jsonl` |  | 200,000 | 319.0 MB |
| `tier1/1.19_multi_move_state_tracking.jsonl` | Apply two to three moves and report the resulting FEN | 160,000 | 390.7 MB |
| `tier1/1.1_fen_to_board.jsonl` | Render a FEN string as a human-readable board diagram | 320,000 | 242.9 MB |
| `tier1/1.2_board_to_fen.jsonl` | Convert a board diagram back to FEN notation | 320,000 | 314.3 MB |
| `tier1/1.3_piece_identification.jsonl` | Identify which piece occupies a given square | 400,000 | 359.3 MB |
| `tier1/1.4_piece_counting.jsonl` | Count pieces of a specific type/color on the board | 320,000 | 516.2 MB |
| `tier1/1.5_state_tracking.jsonl` | Apply moves and report the resulting position | 400,000 | 656.0 MB |
| `tier1/1.6_square_lookup.jsonl` | Read a single square's contents from FEN | 400,000 | 327.6 MB |
| `tier1/1.7_rank_lookup.jsonl` | Read one compressed rank row from FEN | 320,000 | 257.8 MB |
| `tier1/1.8_move_square_edits.jsonl` | Trace square lookups and rank edits for one move | 400,000 | 659.2 MB |
| `tier1/1.9_fen_assembly.jsonl` | Apply one move and assemble the resulting full FEN | 400,000 | 822.1 MB |

### Tier 2 - Rules

| File | Task | Examples | Size |
|------|------|----------|------|
| `tier2/2.0_side_piece_inventory.jsonl` | List side-to-move pieces and squares before move generation | 240,000 | 374.4 MB |
| `tier2/2.10_ray_walk.jsonl` | Walk each slider ray square by square to derive its moves | 200,000 | 305.5 MB |
| `tier2/2.11_legal_filter_trace.jsonl` | Filter every piece's pseudo-legal moves into rejected and legal moves | 160,000 | 445.3 MB |
| `tier2/2.12_illegal_move_correction.jsonl` | Recover from a rejected illegal move by choosing a legal one | 160,000 | 155.4 MB |
| `tier2/2.13_check_evasion.jsonl` | Enumerate every legal escape from check: king moves, checker captures, blocks | 160,000 | 196.7 MB |
| `tier2/2.1_legal_move_gen.jsonl` | List all legal moves for the side to move | 320,000 | 493.4 MB |
| `tier2/2.2_piece_specific_moves.jsonl` | List legal moves for a specific piece | 320,000 | 258.0 MB |
| `tier2/2.3_move_legality_check.jsonl` | Determine whether a move is legal | 320,000 | 271.2 MB |
| `tier2/2.4_check_detection.jsonl` | Detect check, checkmate, or stalemate | 240,000 | 183.5 MB |
| `tier2/2.5_special_rules.jsonl` | Handle castling, en passant, promotion, and 50-move rule | 200,000 | 171.6 MB |
| `tier2/2.6_piece_pseudo_legal_moves.jsonl` |  | 240,000 | 234.9 MB |
| `tier2/2.7_piece_legal_filter.jsonl` |  | 240,000 | 325.1 MB |
| `tier2/2.8_king_safety_filter.jsonl` |  | 240,000 | 262.5 MB |
| `tier2/2.9_legal_moves_by_piece.jsonl` |  | 240,000 | 621.5 MB |

### Tier 3 - Tactics

| File | Task | Examples | Size |
|------|------|----------|------|
| `tier3/3.1_available_captures.jsonl` | Find available capture moves | 240,000 | 159.4 MB |
| `tier3/3.2_threats.jsonl` | Identify pieces that are threatening enemy pieces | 200,000 | 133.5 MB |
| `tier3/3.3_attacked_defended.jsonl` | Count attackers and defenders of a queried square | 240,000 | 290.0 MB |
| `tier3/3.4_tactical_patterns.jsonl` | Recognize tactical motifs | 200,000 | 161.9 MB |
| `tier3/3.5_hanging_pieces.jsonl` | Find undefended pieces that can be captured | 160,000 | 114.2 MB |
| `tier3/3.6_hanging_piece_status.jsonl` | Classify whether one piece is attacked, defended, and hanging | 200,000 | 205.2 MB |
| `tier3/3.7_hanging_piece_filter.jsonl` | Audit attacked pieces and filter defended decoys from hanging pieces | 200,000 | 247.2 MB |
| `tier3/3.8_hanging_piece_claim_verification.jsonl` | Verify and correct hanging-piece claims | 200,000 | 268.8 MB |
| `tier3/3.9_static_exchange_evaluation.jsonl` | Work out capture recapture sequence and net material | 160,000 | 179.1 MB |

### Tier 4 - Evaluation

| File | Task | Examples | Size |
|------|------|----------|------|
| `tier4/4.1_material_balance.jsonl` | Count material and compute the balance | 200,000 | 155.7 MB |
| `tier4/4.2_position_evaluation.jsonl` | Evaluate a position from engine-calibrated labels | 240,000 | 180.7 MB |
| `tier4/4.3_pawn_structure.jsonl` | Analyze pawn structure | 160,000 | 111.5 MB |

### Tier 5 - Openings

| File | Task | Examples | Size |
|------|------|----------|------|
| `tier5/5.1_opening_identification.jsonl` | Name the opening from a position or line | 40,000 | 32.2 MB |
| `tier5/5.2_opening_continuation.jsonl` | Suggest the next book move | 40,000 | 33.8 MB |
| `tier5/5.3_opening_principles.jsonl` | Explain opening principles | 40,000 | 35.1 MB |

### Tier 6 - Endgames

| File | Task | Examples | Size |
|------|------|----------|------|
| `tier6/6.1_endgame_classification.jsonl` | Classify the endgame material | 120,000 | 80.6 MB |
| `tier6/6.2_endgame_wdl.jsonl` | Predict tablebase win/draw/loss | 160,000 | 107.0 MB |
| `tier6/6.3_endgame_best_move.jsonl` | Find a tablebase-backed endgame move | 160,000 | 107.7 MB |
| `tier6/6.4_endgame_principles.jsonl` | Explain endgame principles | 120,000 | 91.5 MB |

### Tier 7 - Planning

| File | Task | Examples | Size |
|------|------|----------|------|
| `tier7/7.10_best_line_trace.jsonl` | Emit a fixed-grammar Stockfish best-line trace | 120,000 | 171.3 MB |
| `tier7/7.11_history_best_move.jsonl` | Imitate the next move from a game prefix and current FEN | 160,000 | 209.0 MB |
| `tier7/7.12_game_episode.jsonl` | Play a multi-turn game window move by move from a FEN anchor | 160,000 | 271.2 MB |
| `tier7/7.13_candidate_compare_trace.jsonl` | Answer a best-move prompt by rating MultiPV candidates inside think before committing | 160,000 | 345.3 MB |
| `tier7/7.14_refute_and_switch.jsonl` | Consider a tempting move, refute it with the engine line, backtrack to the best move | 120,000 | 219.6 MB |
| `tier7/7.15_composed_audit.jsonl` | Audit check/material/captures/hanging inline, then rate candidates and choose | 120,000 | 239.1 MB |
| `tier7/7.16_prose_analysis.jsonl` | Analyze a position in grounded prose (problem/intuition/comparison style), then give the move | 120,000 | 275.1 MB |
| `tier7/7.17_puzzle_episode.jsonl` | Play out a forced puzzle solution turn by turn against injected opponent replies | 160,000 | 184.3 MB |
| `tier7/7.18_line_tracking_episode.jsonl` | State the engine PV as a plan, then execute it across turns tracking the expected replies | 120,000 | 193.4 MB |
| `tier7/7.1_best_move_selection.jsonl` | Select the best move from engine-evaluated positions | 320,000 | 315.9 MB |
| `tier7/7.2_puzzle_solving.jsonl` | Solve a tactical puzzle | 200,000 | 208.2 MB |
| `tier7/7.3_move_consequence.jsonl` | Predict the consequence of a candidate move | 160,000 | 169.0 MB |
| `tier7/7.8_candidate_ratings.jsonl` | Rate five Stockfish MultiPV candidate moves with fixed grammar | 160,000 | 280.2 MB |
| `tier7/7.9_step_verification.jsonl` | Audit a numbered chess trace and identify one broken step | 120,000 | 241.1 MB |