Source 0171a6b7113bc16ed2a0d8b72cebac2ccd96420f. Three actual CUDA FP32 configurations use the same four-RTX5090 host and ordered physical UUID prefixes1/2/4. All share initial M6 weights dcdf28431a7eab40d98f98a1f2ef7b80fbea1b7d9e66f98b6428cfaa67336c85, 7788 eligible protected-filtered selfplay samples e1cc5be4b7d53270631706e3577f9ef4e2376b2ebbd24e4857ddabbf9a45d132, global batch16, seed63001,128 updates,16 warmup updates, AdamW and TF32 disabled. The manifest binds HF input revisions, global order, topology and exact runtime, and includes all three frozen Linux execution scripts with SHA-256. Gradients use rank-ordered FP32 gather/sum/mean, recorded in the v2 binding. The lossless host sample cache is rebuilt after full Rust/protection verification once; every consumer rehashes all feature and row bytes against the identical fixed Samples SHA. No new data generation, architecture comparison, selected stronger model or browser adoption is claimed.

Usage and schema

The measurements split has world_size:int64, primary_json:string, communication_json:string, continuation_json:string and used_gpu_hours_primary:float64. Read pyarrow.parquet.read_table("results.parquet").to_pylist() and decode the JSON columns. Each row is a configuration, not a training sample. Primary throughput covers112 updates/1792 examples after16 warmup updates; max-rank synchronized update sums and post-warmup wall window are separate. Torch peak allocated/reserved bytes are per rank. The isolated gradient-sized all-reduce probe excludes fill/check/barrier and does not measure the communication fraction of training. used_gpu_hours_primary uses active world size times the full128-update attempt wall time; all four physical GPUs remain billed throughout the whole rental, including unused devices, preparation, validation, transfers and retirement.

Verification

Four distinct physical UUIDs each passed an isolated one-visible-device hand matrix/gradient test. The same input SHA and global index schedule were enforced for every world size. Each configuration separately ran64 updates, archived typed JSON/safetensors through the publication harness, downloaded the returned fixed HF revision, restored into a new run/fresh model and AdamW, and completed the remaining64. All six mid/final state archives were byte/hash verified; same-world continuation audits require exact per-rank RNG/cursor and fixed predeclared model/optimizer tolerances. Continuation validation is excluded from primary throughput. Actual deltas, tolerances, precision, timings and memory scopes are in each row. No cross-world bit-exactness, universal scaling claim or later resource-release proof is inferred.

Access and license

Self-generated measurements and metadata, CC-BY-4.0; attribute VoxelShogi contributors and fixed source/HF revisions. Public/manual approval, Community disabled. Models and training states remain separately gated Apache-2.0 and training data CC-BY-4.0. Embedded execution scripts retain Apache-2.0 with their license text in the manifest. No tensors, dataset payloads, provider credentials, HF tokens, raw logs or human games are included.

Downloads last month: -

Total file size:

204 kB