Upload pe_interleaved_alibi_rope.json with huggingface_hub
Browse files
pe_interleaved_alibi_rope.json
ADDED
|
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_class": "BERTModel",
|
| 3 |
+
"model_config": {
|
| 4 |
+
"name": "Albertina-Pos-Encoding_Interleaved_alibi_rope",
|
| 5 |
+
"hidden_size": 768,
|
| 6 |
+
"ffn_factor": 3.0,
|
| 7 |
+
"vocab_size": 32768,
|
| 8 |
+
"bos_token_id": 5,
|
| 9 |
+
"eos_token_id": 6,
|
| 10 |
+
"pad_token_id": 0,
|
| 11 |
+
"mask_token_id": 4,
|
| 12 |
+
"masked_substitution_rate": 0.15,
|
| 13 |
+
"num_hidden_layers": 12,
|
| 14 |
+
"num_attention_heads": 12,
|
| 15 |
+
"tie_word_embeddings": false,
|
| 16 |
+
"rms_norm_eps": 1e-6,
|
| 17 |
+
"attention_type": [],
|
| 18 |
+
"max_position_embeddings": 1024,
|
| 19 |
+
"block_size_for_attention": 128,
|
| 20 |
+
"rope_theta": 10000.0,
|
| 21 |
+
"compile_flexattn": false,
|
| 22 |
+
"bias": false,
|
| 23 |
+
"training_objective": "masked",
|
| 24 |
+
"is_causal": false,
|
| 25 |
+
"default_layer": {
|
| 26 |
+
"attn_impl": "flash",
|
| 27 |
+
"sliding_window_size": null,
|
| 28 |
+
"positional_encoding": "alibi",
|
| 29 |
+
"normalization": "rmsnorm",
|
| 30 |
+
"normalization_position": "pre",
|
| 31 |
+
"ffn_activation": "swiglu",
|
| 32 |
+
"hooks": {}
|
| 33 |
+
},
|
| 34 |
+
"custom_layers": {
|
| 35 |
+
"odd": {
|
| 36 |
+
"attn_impl": "flash",
|
| 37 |
+
"sliding_window_size": null,
|
| 38 |
+
"positional_encoding": "alibi",
|
| 39 |
+
"normalization": "rmsnorm",
|
| 40 |
+
"normalization_position": "pre",
|
| 41 |
+
"ffn_activation": "swiglu",
|
| 42 |
+
"hooks": {}
|
| 43 |
+
},
|
| 44 |
+
"even": {
|
| 45 |
+
"attn_impl": "flash",
|
| 46 |
+
"sliding_window_size": null,
|
| 47 |
+
"positional_encoding": "rope",
|
| 48 |
+
"normalization": "rmsnorm",
|
| 49 |
+
"normalization_position": "pre",
|
| 50 |
+
"ffn_activation": "swiglu",
|
| 51 |
+
"hooks": {}
|
| 52 |
+
}
|
| 53 |
+
}
|
| 54 |
+
},
|
| 55 |
+
"training": {
|
| 56 |
+
"optimizer": "muon",
|
| 57 |
+
"lr_scheduling": true,
|
| 58 |
+
"lr": 8e-4,
|
| 59 |
+
"final_lr": 4e-5,
|
| 60 |
+
"hold_steps": 0.31,
|
| 61 |
+
"weight_decay": 0.01,
|
| 62 |
+
"scheduler": "custom",
|
| 63 |
+
"gradient_clip_val": 1.0,
|
| 64 |
+
"warmup_steps": 0.09,
|
| 65 |
+
"max_epochs": 1,
|
| 66 |
+
"accumulate_grad_batches": 5,
|
| 67 |
+
"seed": 27,
|
| 68 |
+
"save_every_n_steps": 100000,
|
| 69 |
+
"checkpoint_name": "albertina_pos_encoding_interleaved_alibi_rope"
|
| 70 |
+
},
|
| 71 |
+
"tokenizer": {
|
| 72 |
+
"type": "huggingface",
|
| 73 |
+
"pretrained_name": "mrinaldi/Gettone",
|
| 74 |
+
"varlen_strategy": "unpadding"
|
| 75 |
+
},
|
| 76 |
+
"data": {
|
| 77 |
+
"data_root": "/home/matteo/Albertone/Albertina/micro_albertina",
|
| 78 |
+
"batch_size": 56,
|
| 79 |
+
"num_workers": 1,
|
| 80 |
+
"mdat_strategy": "Gettone_1024",
|
| 81 |
+
"mdat_view": null
|
| 82 |
+
},
|
| 83 |
+
"save_dir": "./checkpoints_pos_enc",
|
| 84 |
+
"wandb_project": "Albertina-Pos-Encoding",
|
| 85 |
+
"wandb_run_name": "Albertina-Pos-Encoding_Interleaved_alibi_rope"
|
| 86 |
+
}
|