{ "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": 1, "direction": "latin->khmer", "init": null, "released_as": "samdei/samdei-type-base-pretrained", "train": [ { "file": "generator_v1/train_web_s1.parquet", "repeat": 1 }, { "file": "generator_v1/train_web_s2.parquet", "repeat": 1 }, { "file": "llm/train_gemma_s1.parquet", "repeat": 1 } ], "monitor": "generator_v1/train_web_s1.dev.parquet", "steps": 60000, "selected_step": 60000, "warmup": 3000, "peak_lr": 0.0007, "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-mini", "stage": 1, "direction": "latin->khmer", "init": null, "released_as": "samdei/samdei-type-mini-pretrained", "train": [ { "file": "generator_v2/train_v2_s3.parquet", "repeat": 1 }, { "file": "llm/train_gemrom.parquet", "repeat": 3 } ], "monitor": "generator_v2/train_v2_s3.dev.parquet", "steps": 40000, "selected_step": 35000, "warmup": 2000, "peak_lr": 0.001, "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-base: channel model", "stage": 1, "direction": "khmer->latin", "init": null, "released_as": "samdei/samdei-type-base: channel/", "train": [ { "file": "generator_v1/train_web_s1.parquet", "repeat": 1 }, { "file": "llm/train_gemma_s1.parquet", "repeat": 1 }, { "file": "llm/train_gemrom.parquet", "repeat": 1 } ], "monitor": "generator_v1/train_web_s1.dev.parquet", "steps": 25000, "selected_step": 20000, "warmup": 2000, "peak_lr": 0.001, "max_tokens_per_batch": 12000, "seed": 1, "selected_on": "lowest greedy CER on the monitor file", "model_config": { "d_model": 256, "nhead": 4, "enc_layers": 6, "dec_layers": 2, "ff": 1024, "dropout": 0.1, "max_len": 256, "src_vocab": 228, "tgt_vocab": 216 } } ] }