---
license: mit
pretty_name: Chess master v1
task_categories:
- other
tags:
- chess
- stockfish
- soft-labels
- tablebase
- puzzles
- multipv
configs:
- config_name: positions
  data_files: positions/*.parquet
- config_name: annotations
  data_files: annotations/*.parquet
- config_name: membership
  data_files: membership/*.parquet
- config_name: quarantine
  data_files: quarantine/*.parquet
---

# avewright/chess-master-v1

Linked master chess tables: **positions**, **annotations**, and **membership**.
One position can keep SF19, Lichess MultiPV, a puzzle line, Syzygy, and a model
mistake at the same time. Flat training mixes are recipe exports, not the source
of truth.

Join on `position_id`. Absent metadata is `null` / `unknown`, never coerced to 0
or "verified".

## Load

```python
from datasets import load_dataset
positions = load_dataset("avewright/chess-master-v1", "positions", split="train")
annotations = load_dataset("avewright/chess-master-v1", "annotations", split="train")
membership = load_dataset("avewright/chess-master-v1", "membership", split="train")
```

## Counts

| Table | Rows |
|---|---|
| Mix positions / annotations / membership | 1,004,000 each |
| Unique `position_id` | 1,004,293 |
| Train ∩ eval | 0 |
| Value-eligible | 451,064 (SF19 only) |
| Depth sentinels | 72,565 |
| Disagreeing best-move positions | 66 |

See `quality_report.json` and `SCHEMA.md`.

## Label contracts

- SF19 CP/mate/WDL are White-absolute. Lichess perspective is unverified.
- `soft_probs` are **probabilities**, not logits. The transform is recorded.
- Puzzle FEN is before the opponent setup move. Keep the full line. No invented CP/WDL.
- Syzygy DTZ is not mate. Value stays masked until conversion is verified.
- Off-PV model moves have unknown regret, not an assumed penalty.
- Recipes must not silently relax quality gates.

## Pinned sources

| Source | Revision | License |
|---|---|---|
| `avewright/chess-soft-sf19` | `68cef6c9ba62c62f904f25305f1a7489dab825e0` | MIT |
| `avewright/chess-soft-multipv-lichess` | `291acea2a2ce6fc2aba4b289110d1920cf969afe` | MIT |
| `avewright/chess-soft-syzygy` | `3889985c482837aa3082e182d8f123e005e2c0d4` | MIT |
| `Lichess/chess-puzzles` | `479ea9bc9f681385f5adb23fa27a96c2dc8ae599` | CC0-1.0 |
| `avewright/chess-soft-100m-swa-mistakes` | `69b75e12ab2867c8b9a63d385065dfe646cf86e1` | MIT |

Baseline recipe: `recipes/pilot_45_35_15_5.json` (45% SF19 / 35% Lichess / 15% puzzles / 5% Syzygy).
That mix is a baseline, not an optimum.

This pack is **not** a claim that eval rows are unseen by older checkpoints.