Upload folder using huggingface_hub
Browse files- ckpt.pth +3 -0
- config.json +135 -0
ckpt.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:36441afd84ee5c1abb18a766c4039ebbac71ceb050054e7038d1fe7438dbbea7
|
| 3 |
+
size 87719689
|
config.json
ADDED
|
@@ -0,0 +1,135 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"Qwen3ForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"args_ours": {
|
| 6 |
+
"RQVAE_path": null,
|
| 7 |
+
"alpha_codebook": null,
|
| 8 |
+
"alpha_contrastive": 0.0,
|
| 9 |
+
"alpha_kl": null,
|
| 10 |
+
"alpha_masked_kl": 0.0,
|
| 11 |
+
"alpha_multi_task": 0.0,
|
| 12 |
+
"alpha_nll": 1.0,
|
| 13 |
+
"alpha_ortho": null,
|
| 14 |
+
"alpha_paraphrase": 0.0,
|
| 15 |
+
"alpha_rqvae": null,
|
| 16 |
+
"answers_path": null,
|
| 17 |
+
"assist_content_kl": "content",
|
| 18 |
+
"assist_content_nll": "content",
|
| 19 |
+
"base_model": null,
|
| 20 |
+
"built_in_rqvae": null,
|
| 21 |
+
"chat_format": "icae",
|
| 22 |
+
"checkpoint_path": null,
|
| 23 |
+
"checkpointing_steps": 1000,
|
| 24 |
+
"clip_grad_norm": -1.0,
|
| 25 |
+
"codebook_force": null,
|
| 26 |
+
"completed_steps": null,
|
| 27 |
+
"config": "config/language_modeling/pretrain_icae_qwen.yaml",
|
| 28 |
+
"ctx_change": "gt",
|
| 29 |
+
"ctx_change_ratio": 2.0,
|
| 30 |
+
"ctx_change_shuffle_ratio": 2.0,
|
| 31 |
+
"ctx_kl_weight": null,
|
| 32 |
+
"ctx_kl_weight_group": [],
|
| 33 |
+
"ctx_nll": "gt",
|
| 34 |
+
"ctx_nll_gt_prob": 0.0,
|
| 35 |
+
"ctx_nll_weight": null,
|
| 36 |
+
"ctx_nll_weight_group": [],
|
| 37 |
+
"ctx_select": null,
|
| 38 |
+
"ctx_student": "gt",
|
| 39 |
+
"ctx_student_gt_prob": 0.0,
|
| 40 |
+
"ctx_teacher": "gt",
|
| 41 |
+
"ctx_teacher_gt_prob": 0.0,
|
| 42 |
+
"demo": null,
|
| 43 |
+
"dev_file": "./data/pretrain/dev.jsonl",
|
| 44 |
+
"distill_topk": null,
|
| 45 |
+
"eval_file": null,
|
| 46 |
+
"exclude_dataset_type": null,
|
| 47 |
+
"exp_name": "pretrain_icae_16",
|
| 48 |
+
"exp_note": null,
|
| 49 |
+
"gradient_accumulation_steps": 12,
|
| 50 |
+
"gradient_checkpointing": null,
|
| 51 |
+
"icae_freeze": false,
|
| 52 |
+
"icae_mem_size": 16,
|
| 53 |
+
"icae_name_or_path": "Qwen/Qwen3-8B",
|
| 54 |
+
"icae_projector": null,
|
| 55 |
+
"index_path": null,
|
| 56 |
+
"init_codebook": null,
|
| 57 |
+
"kl_exclude_first_token": false,
|
| 58 |
+
"kl_temperature": null,
|
| 59 |
+
"kl_weight": 2.0,
|
| 60 |
+
"learning_rate": 0.0002,
|
| 61 |
+
"logging_steps": 1,
|
| 62 |
+
"lora_param_version": "v4",
|
| 63 |
+
"lr_scheduler_type": "linear",
|
| 64 |
+
"max_seq_length": 336,
|
| 65 |
+
"max_train_samples": 2000000,
|
| 66 |
+
"max_train_steps": 5209,
|
| 67 |
+
"model_name_or_path": "Qwen/Qwen3-8B",
|
| 68 |
+
"multiple_projector": null,
|
| 69 |
+
"nll_weight": 2.0,
|
| 70 |
+
"num_chunking": 1,
|
| 71 |
+
"num_train_epochs": 1,
|
| 72 |
+
"num_views": 1,
|
| 73 |
+
"only_first_token": null,
|
| 74 |
+
"output_dir": "./wandb/run-20250628_093437-4p0ubi4c/files/checkpoint",
|
| 75 |
+
"overwrite_cache": false,
|
| 76 |
+
"per_device_train_batch_size": 8,
|
| 77 |
+
"preprocessing_chat_format": "icae",
|
| 78 |
+
"preprocessing_num_workers": 16,
|
| 79 |
+
"prob_masked_kl": 0.0,
|
| 80 |
+
"project_name": "icae_pretraining",
|
| 81 |
+
"refer_context": null,
|
| 82 |
+
"ret_embedding_path": null,
|
| 83 |
+
"retrieval_context_length": 180,
|
| 84 |
+
"retriever_name_or_path": null,
|
| 85 |
+
"same_nll_kl": true,
|
| 86 |
+
"seed": 980406,
|
| 87 |
+
"select_criteria": null,
|
| 88 |
+
"select_gt_ratio": 0.0,
|
| 89 |
+
"select_negative_ratio": 0.0,
|
| 90 |
+
"sum_codebook": null,
|
| 91 |
+
"task_type": "pretrain",
|
| 92 |
+
"train_file": "./data/pretrain/train.jsonl",
|
| 93 |
+
"train_shuffle": true,
|
| 94 |
+
"update_codebook": null,
|
| 95 |
+
"update_projector_only": true,
|
| 96 |
+
"use_fast_tokenizer": null,
|
| 97 |
+
"use_flash_attn": true,
|
| 98 |
+
"use_rag_tuning": null,
|
| 99 |
+
"use_xrag_embedding": null,
|
| 100 |
+
"warmup_ratio": 0.03,
|
| 101 |
+
"weight_decay": 0.0,
|
| 102 |
+
"workdir": "."
|
| 103 |
+
},
|
| 104 |
+
"attention_bias": false,
|
| 105 |
+
"attention_dropout": 0.0,
|
| 106 |
+
"bos_token_id": 151643,
|
| 107 |
+
"eos_token_id": 151645,
|
| 108 |
+
"head_dim": 128,
|
| 109 |
+
"hidden_act": "silu",
|
| 110 |
+
"hidden_size": 4096,
|
| 111 |
+
"icae_projector": null,
|
| 112 |
+
"initializer_range": 0.02,
|
| 113 |
+
"intermediate_size": 12288,
|
| 114 |
+
"max_position_embeddings": 40960,
|
| 115 |
+
"max_window_layers": 36,
|
| 116 |
+
"mem_size": 16,
|
| 117 |
+
"model_type": "qwen3",
|
| 118 |
+
"multiple_projector": null,
|
| 119 |
+
"num_attention_heads": 32,
|
| 120 |
+
"num_hidden_layers": 36,
|
| 121 |
+
"num_key_value_heads": 8,
|
| 122 |
+
"projector_type": "mlp2x_gelu",
|
| 123 |
+
"retrieval_embed_length": 16,
|
| 124 |
+
"retriever_hidden_size": 4096,
|
| 125 |
+
"rms_norm_eps": 1e-06,
|
| 126 |
+
"rope_scaling": null,
|
| 127 |
+
"rope_theta": 1000000,
|
| 128 |
+
"sliding_window": null,
|
| 129 |
+
"tie_word_embeddings": false,
|
| 130 |
+
"torch_dtype": "bfloat16",
|
| 131 |
+
"transformers_version": "4.51.0",
|
| 132 |
+
"use_cache": true,
|
| 133 |
+
"use_sliding_window": false,
|
| 134 |
+
"vocab_size": 151936
|
| 135 |
+
}
|