jbduran's picture
Import experiment archive from bart (batch 2)
6951e66 verified
Raw
History Blame Contribute Delete
2.78 kB
{
"schema_version": 1,
"stage": "base",
"experiment_id": "Think.Unbounded-d32",
"datasets": {
"original": {
"adapter": "parquet_shards",
"repo": "jbduran/think-dataset-clean-1930s",
"revision": "45225d95bc15f942be3b4b344738cea1f66e3de8",
"validation_shard": 472,
"num_train_shards": 330,
"download_workers": 4
},
"midtrain_r30": {
"adapter": "parquet_shards",
"repo": "zachnorton03/think-midtrain",
"revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
"subfolder": "mixed/ratio_30/data",
"validation_shard": 1262,
"num_train_shards": 391,
"download_workers": 4
},
"midtrain_r60": {
"adapter": "parquet_shards",
"repo": "zachnorton03/think-midtrain",
"revision": "9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc",
"subfolder": "mixed/ratio_60/data",
"validation_shard": 523,
"num_train_shards": 207,
"download_workers": 4
}
},
"mixture_schedule": {
"total_tokens": 20132659200,
"seed_data": 1234,
"max_epochs": 4,
"stages": [
{
"name": "base",
"start_tokens": 0,
"source": "original"
},
{
"name": "injection",
"start_tokens": 14092861440,
"source": "midtrain_r30"
},
{
"name": "decay_mix",
"start_tokens": 18119393280,
"source": "midtrain_r60"
}
]
},
"tokenizer": {
"mode": "reuse",
"source_experiment_id": "clean1930s-d24-r12-ctx4096-fulltok-v1",
"vocab_size": 32768
},
"pretokenize": {
"enabled": true,
"slack": 1.03,
"require_no_wrap": true,
"val_tokens": 20971520,
"shard_tokens": 100000000,
"tokenizer_threads": 8
},
"training": {
"depth": 32,
"scaling_params": 1677724672,
"target_param_data_ratio": 12,
"max_seq_len": 4096,
"window_pattern": "SSSL",
"muon_momentum": 0.9,
"device_batch_size": 2,
"total_batch_size": 2097152,
"fp8": true,
"fp8_recipe": "tensorwise",
"seed": 42,
"save_every": 500,
"eval_every": 250,
"eval_tokens": 2097152,
"core_metric_every": -1,
"sample_every": -1
},
"artifacts": {
"repo": "jbduran/think.nano",
"keep_local_checkpoints": 1
},
"wandb": {
"entity": "jbduran-thinkingmachinesncsu",
"project": "think.nano",
"name": "Think.Unbounded-d32",
"group": "clean1930s-d32",
"tags": [
"think-dataset-clean-1930s",
"d32",
"ratio12",
"ctx4096",
"sssl",
"fp8",
"muon-momentum-constant",
"muon-momentum-0.90",
"midtrain-schedule",
"mix-og",
"0-30-60"
]
},
"config_fingerprint": "6f8ae461e0eb9ee2",
"artifact_path": "experiments/Think.Unbounded-d32"
}