---
license: mit
task_categories:
- other
tags:
- chess
- stockfish
- multipv
- soft-targets
- policy-distillation
size_categories:
- 1M<n<10M
---

# exp186 Stockfish MultiPV soft targets (~2M)

Local Stockfish 18 full-strength MultiPV harvest for chess transformer soft-policy training.

## Contents
- `data/positions_*.jsonl` — ~2,000,000 positions (400 shards)
- `soft_cache.pt` — training-ready tensor cache (compact move vocab 1968)

## Labeling
- Engine: Stockfish 18 (full strength, no UCI_LimitStrength)
- MultiPV: 8
- Depths: sampled in [2, 8] (weighted toward mid depths)
- Soft probs: `softmax(cp / τ)` with τ=120

## Schema
Each JSONL line has `fen`, `best_move`, `best_cp`, `soft_targets` (up to 8 moves with `uci`/`prob`/`cp`), `label_depth`, `phase`, `source`.

## Stats
- written: 1998784
- depth_hist: `{"2": 209465, "3": 315864, "4": 421147, "5": 314654, "6": 210726, "7": 104881, "8": 105176}`
- uploaded_at: 2026-07-10T01:59:56.219054+00:00