kafka-duplex-stage1-trainclean460 / stage1_trainclean460.json
Praha-Labs's picture
Upload stage1_trainclean460.json with huggingface_hub
8d968fe verified
Raw
History Blame Contribute Delete
1.01 kB
{
"stage": "stage1_alignment",
"dataset": {
"name": "LibriSpeechTrainClean460Bootstrap",
"train_manifest": "/workspace/kafka_duplex_data/stage1_trainclean460/train.stage1.jsonl",
"val_manifest": "/workspace/kafka_duplex_data/stage1_trainclean460/val.stage1.jsonl",
"text_tokenizer": "mock_text_ids_until_real_tokenizer_is_wired",
"speech_codec": "mock",
"chunk_ms": 200,
"speech_tokens_per_chunk": 10
},
"model": {
"backbone": "Stage1BootstrapLM",
"context_length": 1024,
"vocab_size": 53260,
"hidden_size": 768,
"num_layers": 12,
"num_heads": 12,
"ffw_multiplier": 4,
"dropout": 0.1
},
"optimization": {
"epochs": 3,
"effective_batch_size": 32,
"micro_batch_size": 4,
"learning_rate": 0.0001,
"warmup_steps": 1000,
"weight_decay": 0.01,
"max_grad_norm": 1.0
},
"evaluation": {
"eval_every_steps": 1000,
"checkpoint_every_steps": 5000,
"wer_num_samples": 64,
"tts_num_prompts": 8
}
}