Hugging Face's logo

neginashz
/
qwen-qlora-fsdp

PEFT
Safetensors
qwen2
axolotl
Generated from Trainer
4-bit precision
bitsandbytes
Model card Files Community
neginashz commited on
Commit
05be873
·
verified ·
1 Parent(s): a299aa3

Model save

Browse files
Files changed (1) hide show
  1. README.md +164 -0
README.md ADDED
@@ -0,0 +1,164 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: peft
3
+ license: apache-2.0
4
+ base_model: Qwen/Qwen2.5-7B-Instruct
5
+ tags:
6
+ - axolotl
7
+ - generated_from_trainer
8
+ datasets:
9
+ - medalpaca/medical_meadow_medqa
10
+ model-index:
11
+ - name: qwen-lora-fsdp
12
+ results: []
13
+ ---
14
+
15
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
16
+ should probably proofread and complete it, then remove this comment. -->
17
+
18
+ [<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
19
+ <details><summary>See axolotl config</summary>
20
+
21
+ axolotl version: `0.6.0`
22
+ ```yaml
23
+ base_model: Qwen/Qwen2.5-7B-Instruct
24
+ trust_remote_code: true
25
+
26
+ load_in_8bit: false
27
+ load_in_4bit: true
28
+ strict: false
29
+
30
+ datasets:
31
+ - path: medalpaca/medical_meadow_medqa
32
+ type: alpaca
33
+ dataset_prepared_path:
34
+ val_set_size: 0.2
35
+ output_dir: ./outputs/out
36
+
37
+ sequence_len: 2048
38
+ sample_packing: true
39
+ eval_sample_packing: true
40
+ pad_to_sequence_len: true
41
+
42
+ adapter: qlora
43
+ lora_model_dir:
44
+ lora_r: 256
45
+ lora_alpha: 128
46
+ lora_dropout: 0.05
47
+ lora_target_linear: true
48
+ lora_fan_in_fan_out:
49
+
50
+ wandb_project:
51
+ wandb_entity:
52
+ wandb_watch:
53
+ wandb_name:
54
+ wandb_log_model:
55
+
56
+ gradient_accumulation_steps: 1
57
+ micro_batch_size: 1
58
+ num_epochs: 3
59
+ optimizer: adamw_torch
60
+ lr_scheduler: cosine
61
+ learning_rate: 0.00002
62
+
63
+ train_on_inputs: false
64
+ group_by_length: false
65
+ bf16: true
66
+ fp16:
67
+ tf32:
68
+
69
+ gradient_checkpointing: true
70
+ gradient_checkpointing_kwargs:
71
+ use_reentrant: false
72
+ early_stopping_patience:
73
+ resume_from_checkpoint:
74
+ local_rank:
75
+ logging_steps: 10
76
+ xformers_attention:
77
+ flash_attention: true
78
+
79
+ warmup_steps:
80
+ evals_per_epoch: 4
81
+ saves_per_epoch: 1
82
+ debug:
83
+ deepspeed:
84
+ weight_decay: 0.0
85
+ fsdp:
86
+ - full_shard
87
+ - auto_wrap
88
+ fsdp_config:
89
+ fsdp_limit_all_gathers: true
90
+ fsdp_sync_module_states: true
91
+ fsdp_offload_params: true
92
+ fsdp_use_orig_params: false
93
+ fsdp_cpu_ram_efficient_loading: true
94
+ fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
95
+ fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
96
+ fsdp_state_dict_type: FULL_STATE_DICT
97
+ fsdp_sharding_strategy: FULL_SHARD
98
+ special_tokens:
99
+
100
+ hub_model_id: neginashz/qwen-lora-fsdp
101
+ ```
102
+
103
+ </details><br>
104
+
105
+ # qwen-lora-fsdp
106
+
107
+ This model is a fine-tuned version of [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) on the medalpaca/medical_meadow_medqa dataset.
108
+ It achieves the following results on the evaluation set:
109
+ - Loss: 0.1946
110
+
111
+ ## Model description
112
+
113
+ More information needed
114
+
115
+ ## Intended uses & limitations
116
+
117
+ More information needed
118
+
119
+ ## Training and evaluation data
120
+
121
+ More information needed
122
+
123
+ ## Training procedure
124
+
125
+ ### Training hyperparameters
126
+
127
+ The following hyperparameters were used during training:
128
+ - learning_rate: 2e-05
129
+ - train_batch_size: 1
130
+ - eval_batch_size: 1
131
+ - seed: 42
132
+ - distributed_type: multi-GPU
133
+ - num_devices: 2
134
+ - total_train_batch_size: 2
135
+ - total_eval_batch_size: 2
136
+ - optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
137
+ - lr_scheduler_type: cosine
138
+ - lr_scheduler_warmup_steps: 55
139
+ - num_epochs: 3
140
+
141
+ ### Training results
142
+
143
+ | Training Loss | Epoch | Step | Validation Loss |
144
+ |:-------------:|:------:|:----:|:---------------:|
145
+ | 0.1208 | 0.2504 | 154 | 0.1264 |
146
+ | 0.1361 | 0.5008 | 308 | 0.1208 |
147
+ | 0.1211 | 0.7512 | 462 | 0.1133 |
148
+ | 0.1154 | 1.0016 | 616 | 0.1119 |
149
+ | 0.076 | 1.2504 | 770 | 0.1208 |
150
+ | 0.0563 | 1.5008 | 924 | 0.1302 |
151
+ | 0.0493 | 1.7512 | 1078 | 0.1320 |
152
+ | 0.0678 | 2.0016 | 1232 | 0.1255 |
153
+ | 0.0091 | 2.2504 | 1386 | 0.1796 |
154
+ | 0.0216 | 2.5008 | 1540 | 0.1903 |
155
+ | 0.0105 | 2.7512 | 1694 | 0.1946 |
156
+
157
+
158
+ ### Framework versions
159
+
160
+ - PEFT 0.14.0
161
+ - Transformers 4.47.0
162
+ - Pytorch 2.5.1+cu124
163
+ - Datasets 3.1.0
164
+ - Tokenizers 0.21.0