mengsay's picture
Link the per-size pretrained repos (samdei-type-mini-pretrained, samdei-type-base-pretrained)
d21bb7a verified
Raw History Blame Contribute Delete
5.23 kB
{
"about": "Exact training mixes of the Samdei Type checkpoints trained on this dataset, read from the arguments stored in each checkpoint (scripts/train.py). Files are listed in the order given to the training script; 'repeat' is how many times the file was listed. Paths are relative to this dataset unless they name another repo. 'monitor' is the synthetic held-out file watched during training; checkpoints were selected as given in 'selected_on'.",
"shared": {
"optimizer": "AdamW, betas 0.9/0.98, weight decay 0.01",
"schedule": "cosine decay to 5% of the peak",
"label_smoothing": 0.1,
"precision": "bf16 autocast",
"batching": "about 12k tokens per batch, 2-D length buckets padded to multiples of 16"
},
"runs": [
{
"model": "samdei-type-base",
"stage": 2,
"direction": "latin->khmer",
"init": "samdei/samdei-type-base-pretrained",
"released_as": null,
"train": [
{
"file": "generator/ft_v2.parquet",
"repeat": 1
},
{
"file": "abbreviated/ft_v2_abbr3.parquet",
"repeat": 1
},
{
"file": "llm_romanized/train_gemrom2.parquet",
"repeat": 4
},
{
"file": "abbreviated/train_gemrom2_abbr3.parquet",
"repeat": 2
},
{
"file": "mined_tables/ft_rules_chhwd.parquet",
"repeat": 1
}
],
"monitor": "validation/train_v2_s3.dev.parquet",
"steps": 6000,
"selected_step": 4000,
"warmup": 300,
"peak_lr": 0.0002,
"max_tokens_per_batch": 12000,
"seed": 1,
"selected_on": "lowest mean greedy CER on Dev-Sentences and Akkhara-Preview (selection only; never trained on)",
"model_config": {
"d_model": 384,
"nhead": 6,
"enc_layers": 6,
"dec_layers": 3,
"ff": 1536,
"dropout": 0.1,
"max_len": 256,
"src_vocab": 216,
"tgt_vocab": 228
}
},
{
"model": "samdei-type-base",
"stage": 3,
"direction": "latin->khmer",
"init": "Base stage 2",
"released_as": "samdei/samdei-type-base: model.safetensors (and onnx/)",
"train": [
{
"file": "generator/ft_v2.parquet",
"repeat": 1
},
{
"file": "abbreviated/ft_v2_abbr3.parquet",
"repeat": 1
},
{
"file": "llm_romanized/train_gemrom2.parquet",
"repeat": 4
},
{
"file": "abbreviated/train_gemrom2_abbr3.parquet",
"repeat": 2
},
{
"file": "mined_tables/ft_rules_chhwd.parquet",
"repeat": 1
},
{
"file": "casual/ft_casual_noakk.parquet",
"repeat": 1
}
],
"monitor": "validation/train_v2_s3.dev.parquet",
"steps": 3000,
"selected_step": 3000,
"warmup": 100,
"peak_lr": 0.0001,
"max_tokens_per_batch": 12000,
"seed": 1,
"selected_on": "lowest greedy CER on the first 400 messages of the closed development set",
"model_config": {
"d_model": 384,
"nhead": 6,
"enc_layers": 6,
"dec_layers": 3,
"ff": 1536,
"dropout": 0.1,
"max_len": 256,
"src_vocab": 216,
"tgt_vocab": 228
}
},
{
"model": "samdei-type-mini",
"stage": 2,
"direction": "latin->khmer",
"init": "samdei/samdei-type-mini-pretrained",
"released_as": null,
"train": [
{
"file": "generator/ft_v2.parquet",
"repeat": 1
},
{
"file": "abbreviated/ft_v2_abbr3.parquet",
"repeat": 1
},
{
"file": "llm_romanized/train_gemrom2.parquet",
"repeat": 4
},
{
"file": "abbreviated/train_gemrom2_abbr3.parquet",
"repeat": 2
},
{
"file": "mined_tables/ft_rules_chh.parquet",
"repeat": 1
}
],
"monitor": "validation/train_v2_s3.dev.parquet",
"steps": 6000,
"selected_step": 4000,
"warmup": 300,
"peak_lr": 0.0002,
"max_tokens_per_batch": 12000,
"seed": 1,
"selected_on": "lowest mean greedy CER on Dev-Sentences and Akkhara-Preview (selection only; never trained on)",
"model_config": {
"d_model": 256,
"nhead": 4,
"enc_layers": 6,
"dec_layers": 2,
"ff": 1024,
"dropout": 0.1,
"max_len": 256,
"src_vocab": 215,
"tgt_vocab": 228
}
},
{
"model": "samdei-type-mini",
"stage": 3,
"direction": "latin->khmer",
"init": "Mini stage 2",
"released_as": "samdei/samdei-type-mini: model.safetensors (and onnx/)",
"train": [
{
"file": "generator/ft_v2.parquet",
"repeat": 1
},
{
"file": "abbreviated/ft_v2_abbr3.parquet",
"repeat": 1
},
{
"file": "llm_romanized/train_gemrom2.parquet",
"repeat": 4
},
{
"file": "abbreviated/train_gemrom2_abbr3.parquet",
"repeat": 2
},
{
"file": "mined_tables/ft_rules_chh.parquet",
"repeat": 1
},
{
"file": "casual/ft_casual_noakk.parquet",
"repeat": 1
}
],
"monitor": "validation/train_v2_s3.dev.parquet",
"steps": 3000,
"selected_step": 2500,
"warmup": 100,
"peak_lr": 0.0001,
"max_tokens_per_batch": 12000,
"seed": 1,
"selected_on": "lowest greedy CER on the first 400 messages of the closed development set",
"model_config": {
"d_model": 256,
"nhead": 4,
"enc_layers": 6,
"dec_layers": 2,
"ff": 1024,
"dropout": 0.1,
"max_len": 256,
"src_vocab": 215,
"tgt_vocab": 228
}
}
]
}