README.md

metadata

license: mit
pretty_name: exp190 Phase-Balanced Deep Soft Targets
tags:
  - chess
  - stockfish
  - multipv
  - soft-labels
  - imitation-learning
size_categories:
  - 100K<n<1M

exp190 — Phase-balanced deep MultiPV soft targets

Stockfish 18 MultiPV soft-policy labels with phase quotas so training does not drown in openings. Built for policy-only (no-MCTS) ChessTransformer finetunes such as avewright/chess-transformer-200m-maxelo.

Motivation

Naive HF streaming + shallow MultiPV skews toward openings. Elo without search is decided in middlegame/endgame. This harvest enforces:

Phase	Target share	Label depth
Opening	22%	10–14
Middlegame	48%	12–16
Endgame	30%	14–18

Contents

Path	Description
soft_cache.pt	Torch cache ready for exp189-style training
data/*.jsonl	Raw labeled positions (FEN + MultiPV soft targets)
status.json	Harvest counters at upload time
ARCHITECTURE.md	Design notes for stratified training

soft_cache.pt tensors

board_array, turn, castling, ep_square, move_idx, cp, mate, soft_indices, soft_probs, phase, label_depth.

Phase encoding: 0=opening, 1=middlegame, 2=endgame.

How it was generated

• Engine: Stockfish 18 (x86-64-vnni512), full strength, MultiPV 8
• FEN sources: HF position streams, book playouts, endgame templates, random walks
• Deficit-first producer + hard phase gate at writer
• Syzygy tablebases when available
• Code: experiments/exp190_phase_deep_harvest.py in avewright/transform

Recommended training use

1. Mix with shallow volume (exp186-sf-multipv-2m) at ~30–40% deep / 60–70% shallow
2. Stratify batches by phase (1/3 each) so gradients never bog on openings
3. Optional depth weight: w = label_depth / mean_depth (clip 0.5–1.5)

Related

• Shallow soft: avewright/exp186-sf-multipv-2m
• Model trained with this mix: avewright/chess-transformer-200m-maxelo
• Progress write-up: docs/PROGRESS_2026-07-10.md in the transform repo