You can deploy this model and start earning money today!

BertilBraun/
alphazero-chess

Name: alphazero-chess
Availability: InStock

Views

No views yet

AlphaZero Chess (v4 freeze run)

Project homepage · Hugging Face repository

Model source

Repository: BertilBraun/alphazero-chess
Revision used: d39d5c85d00e8a6a94fac856dc188d5468bdd626
Run name: vast-chess-8gpu-1d-r4
Source stage: vast-chess-8gpu-1d-r4 continuation from generation 150
Training period (run directory): four-day run (~2026-08-17 completion window)
Final checkpoint artifact family: latest.jit.pt / latest.pt

Snapshot

Metric	Value
Network	12 layers, hidden size 112, global-pooling residual context (every second block), policy/value channels `4/2`, value FC 48
Trainer	AdamW, BF16, staged LR (`0.004`→`0.003` at gen 350→`0.002` at gen 550), max grad norm `0.5`, global batch `2048`, local batch `256`
Self-play topology	8× RTX 3060, 64 vCPUs, 188.7 GiB RAM
Evaluation schedule	1200 s cadence, 10 fixed dataset + fixed-checkpoint + fixed Stockfish checks
Inference batch settings	TorchScript inference workers: 2, batch size: 64, outstanding batches per worker: 2
Deployment	Lichess and web inference use TorchScript artifact from this run; repository is refreshed via revision-aware Hugging Face lookups

Full training configuration

The values below are the active values from the four-day run manifest:

Experiment/run metadata

Run: vast-chess-8gpu-1d-r4
TensorBoard run directory: vast-chess-8gpu-1d-r4-late-quality
Stage: continuation from generation 150 (vast-chess-8gpu-1d-r3-generation-150/checkpoint_150.json)
Requires approval: true
Random seed: 20260811

Topology and resources

Trainer topology:
- Device type: cuda
- DDP backend: nccl
- Training ranks/devices: 0..7
- Self-play device fanout: 0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7
- Self-play parallel games per process: 512
Evaluation devices/cycle: GPUs [4, 5, 6, 7]
Host/runtime:
- Runtime image: vastai/pytorch:cuda-13.0.3-auto
- Python: 3.12
- PyTorch: 2.12.1+cu126
- CUDA: 12.6
Hard limits:
- Hourly price: 0.4608888889 USD/hour
- Four-day wall cost estimate: 0.4608888889 × 96h ≈ 44.2 USD (excluding one-time setup)
- Max open files: 49152
- Max host RAM usage: 92%
- Minimum free disk: 10 GiB
- Telemetry interval: 5s

Network architecture

Layers: 12, channels hidden: 112
Policy channels: 4
Value channels: 2
Value FC width: 48
Residual context: global_pooling, placement every_second_block

Training hyperparameters

Global batch size: 2048
Local batch size: 256
Optimizer: adamw
Precision: bfloat16
Compilation: disabled
Max grad norm: 0.5
Staged learning rate:
- Generation 0: 0.004
- Generation 350: 0.003
- Generation 550: 0.002
Replay capacity: constant 1,500,000
Replay maximum capacity: 2,500,000
Maximum policy entries per turn: 60
Replay ratio: 8
Optimizer steps per quantum: 500

Chess self-play search

Full-search schedule:
- Generation 0: 600
- Generation 180: 700
- Generation 250: 800
- Generation 550: 1000
Fast-searches: 150
Full-search probability: 0.25
Parallel searches: 1
Inference workers: 2
Inference batch size: 64
Outstanding batches per worker: 2
Dirichlet: ε=0.25, α=0.3
Exploration constant: 1.5
FPU type/reduction: reduced_parent_value, 0.2
Forced playouts: enabled (1.5)
Retained root-visit fraction: 0.6
Greedy after ply: 80
Start/Final temperature: 1.3 → 0.1
Force-fast-search after ply: 160
Detailed statistics workers: 1
Max game length: 200 plies
Start-position controls:
- true-start probability: 0.5
- candidate-visit mass: 0.85
- min/max candidates: 2 / 3
- absolute root-value clip: 0.8
- minimum remaining plies: 25
- archive positions: 50,000
- archive age max generations: 40

Objective and auxiliary losses

Policy loss weight: 1.0
Value loss weight: 1.0
Root value blend: 0.1
Value discount (staged):
- Generation 0: 0.9985
- Generation 300: 0.996
Auxiliary targets:
- next-policy (ply_offset=1, weight 0.1)
- remaining-game-length (weight 0.1, normalization 400)
Resignation (calibrated, from generation 70):
- false non-loss ceiling: 0.025
- continuation probability: 0.1
- max evidence window: 2000
- candidate threshold: [-0.99, -0.7] with step 0.01
- minimum trigger count: 100
- confidence level: 0.95
- max relaxation per generation: 0.01

Evaluation configuration

Cadence: 1200 s
Job timeout: 1800 s
Shutdown grace: 30 s
Bootstrap samples: 10000
Concurrent jobs: 10
Evaluation dataset:
- Path: /workspace/evaluation-artifacts/chess/chess-stockfish-evaluation-v1.bin
- Source seed: 20260811
- Move sampling temperature: 1.0
Openings:
- Path: /workspace/evaluation-artifacts/chess/chess-stockfish-8moves-v3-openings-v1.json
- Opening count: 50
Engine opponent:
- Stockfish, nodes labels: 10000, match nodes 1000
- Threads: 1, hash 1024 MiB, multiPV=8, softmax temp 0.15
Definitions evaluated:
- fixed-dataset
- previous checkpoint checks (previous-20m, previous-40m, previous-60m)
- stockfish levels 0 through 4
- stockfish fixed nodes 1000

Approximate Elo by search budget

These values are from four-day confirmation matches and are intended as calibrated operating references, not universal engine strength claims:

Search budget	Approximate Elo	95% score interval
64 searches	1,954 (1,912–1,996)	57.75% [51.75%, 63.50%]
1,000 searches	2,554 (2,519–2,591)	57.75% [52.75%, 62.75%]
10,000 searches	2,821 (2,781–2,864)	66.75% [61.50%, 72.00%]
1 second	2,695 (2,660–2,730)	49.25% [44.25%, 54.25%]
5 seconds	unavailable (compute-node loss before confirmation completion)	unavailable

Provenance and reproducibility

Source revision: d39d5c85d00e8a6a94fac856dc188d5468bdd626
Approval hash: 9e0478dbe7a29295cd8ee3019abc687332d7e05d52f68ebae6dd8dcb514f8a9e
Configuration hash is in the run manifest and should be used with the corresponding experiment approval.
Approximate file provenance from manifest:
- Initial model SHA: 0d5a2739ec871ef63e43a29d23f7a87182296539371ed811084bfcd8f1be7bf9
- Initial optimizer SHA: 22fb840dcbbacb747175c23da5c565e778d9aecadc2b505d11802a669dfa6a65
- Initial inference model SHA: fbe498d95c5a9de6c8bdca73bdd0a45fa0c30be8fcd30e838a3eae32f1307c70
- Evaluation dataset SHA: 2e3d843e07a26e17a0c4a36f4913e97a54e48a77456941e6e4f44a774dcaccb5
- Evaluation dataset manifest SHA: 95a59c9844fcaf0b40b3a3219e07c16f2a7697c64a5120c54a9091a0b2c31379

BertilBraun/alphazero-chess