Hugging Face's logo

SystemAdmin123
/
tiny-random-LlamaForCausalLM

Text Generation
Transformers
Safetensors
llama
axolotl
Generated from Trainer
conversational
text-generation-inference
Model card Files
xet
Community
SystemAdmin123 commited on
Commit
300dd0a
·
verified ·
1 Parent(s): 7691d37

End of training

Browse files
Files changed (1) hide show
  1. README.md +37 -30
README.md CHANGED
@@ -20,7 +20,7 @@ should probably proofread and complete it, then remove this comment. -->
20
  axolotl version: `0.6.0`
21
  ```yaml
22
  base_model: trl-internal-testing/tiny-random-LlamaForCausalLM
23
- batch_size: 32
24
  bf16: true
25
  chat_template: tokenizer_default_fallback_alpaca
26
  datasets:
@@ -36,26 +36,26 @@ datasets:
36
  system_prompt: ''
37
  device_map: auto
38
  eval_sample_packing: false
39
- eval_steps: 200
40
  flash_attention: true
41
- gpu_memory_limit: 80GiB
42
  group_by_length: true
43
  hub_model_id: SystemAdmin123/tiny-random-LlamaForCausalLM
44
  hub_strategy: checkpoint
45
  learning_rate: 0.0002
46
  logging_steps: 10
47
  lr_scheduler: cosine
48
- max_steps: 2500
49
- micro_batch_size: 4
50
  model_type: AutoModelForCausalLM
51
  num_epochs: 100
52
  optimizer: adamw_bnb_8bit
53
- output_dir: /root/.sn56/axolotl/outputs/tiny-random-LlamaForCausalLM
54
  pad_to_sequence_len: true
55
  resize_token_embeddings_to_32x: false
56
- sample_packing: false
57
- save_steps: 400
58
- save_total_limit: 1
59
  sequence_len: 2048
60
  tokenizer_type: LlamaTokenizerFast
61
  torch_dtype: bf16
@@ -77,7 +77,7 @@ warmup_ratio: 0.05
77
 
78
  This model is a fine-tuned version of [trl-internal-testing/tiny-random-LlamaForCausalLM](https://huggingface.co/trl-internal-testing/tiny-random-LlamaForCausalLM) on the argilla/databricks-dolly-15k-curated-en dataset.
79
  It achieves the following results on the evaluation set:
80
- - Loss: 8.6989
81
 
82
  ## Model description
83
 
@@ -97,36 +97,43 @@ More information needed
97
 
98
  The following hyperparameters were used during training:
99
  - learning_rate: 0.0002
100
- - train_batch_size: 4
101
- - eval_batch_size: 4
102
  - seed: 42
 
 
 
 
103
  - optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
104
  - lr_scheduler_type: cosine
105
- - lr_scheduler_warmup_steps: 125
106
- - training_steps: 2500
107
 
108
  ### Training results
109
 
110
- | Training Loss | Epoch | Step | Validation Loss |
111
- |:-------------:|:------:|:----:|:---------------:|
112
- | No log | 0.0003 | 1 | 10.3763 |
113
- | 9.7054 | 0.0592 | 200 | 9.6862 |
114
- | 8.9091 | 0.1184 | 400 | 8.9612 |
115
- | 8.7257 | 0.1776 | 600 | 8.7627 |
116
- | 8.7416 | 0.2368 | 800 | 8.7109 |
117
- | 8.5944 | 0.2959 | 1000 | 8.6982 |
118
- | 8.673 | 0.3551 | 1200 | 8.6963 |
119
- | 8.7511 | 0.4143 | 1400 | 8.6972 |
120
- | 8.729 | 0.4735 | 1600 | 8.6961 |
121
- | 8.6325 | 0.5327 | 1800 | 8.6948 |
122
- | 8.6338 | 0.5919 | 2000 | 8.6946 |
123
- | 8.7376 | 0.6511 | 2200 | 8.6954 |
124
- | 8.573 | 0.7103 | 2400 | 8.6989 |
 
 
 
125
 
126
 
127
  ### Framework versions
128
 
129
  - Transformers 4.48.1
130
- - Pytorch 2.4.1+cu124
131
  - Datasets 3.2.0
132
  - Tokenizers 0.21.0
 
20
  axolotl version: `0.6.0`
21
  ```yaml
22
  base_model: trl-internal-testing/tiny-random-LlamaForCausalLM
23
+ batch_size: 64
24
  bf16: true
25
  chat_template: tokenizer_default_fallback_alpaca
26
  datasets:
 
36
  system_prompt: ''
37
  device_map: auto
38
  eval_sample_packing: false
39
+ eval_steps: 40
40
  flash_attention: true
41
+ gradient_checkpointing: true
42
  group_by_length: true
43
  hub_model_id: SystemAdmin123/tiny-random-LlamaForCausalLM
44
  hub_strategy: checkpoint
45
  learning_rate: 0.0002
46
  logging_steps: 10
47
  lr_scheduler: cosine
48
+ max_steps: 5000
49
+ micro_batch_size: 32
50
  model_type: AutoModelForCausalLM
51
  num_epochs: 100
52
  optimizer: adamw_bnb_8bit
53
+ output_dir: /root/.sn56/axolotl/tmp/tiny-random-LlamaForCausalLM
54
  pad_to_sequence_len: true
55
  resize_token_embeddings_to_32x: false
56
+ sample_packing: true
57
+ save_steps: 20
58
+ save_total_limit: 2
59
  sequence_len: 2048
60
  tokenizer_type: LlamaTokenizerFast
61
  torch_dtype: bf16
 
77
 
78
  This model is a fine-tuned version of [trl-internal-testing/tiny-random-LlamaForCausalLM](https://huggingface.co/trl-internal-testing/tiny-random-LlamaForCausalLM) on the argilla/databricks-dolly-15k-curated-en dataset.
79
  It achieves the following results on the evaluation set:
80
+ - Loss: 9.1944
81
 
82
  ## Model description
83
 
 
97
 
98
  The following hyperparameters were used during training:
99
  - learning_rate: 0.0002
100
+ - train_batch_size: 32
101
+ - eval_batch_size: 32
102
  - seed: 42
103
+ - distributed_type: multi-GPU
104
+ - num_devices: 2
105
+ - total_train_batch_size: 64
106
+ - total_eval_batch_size: 64
107
  - optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
108
  - lr_scheduler_type: cosine
109
+ - lr_scheduler_warmup_steps: 30
110
+ - training_steps: 600
111
 
112
  ### Training results
113
 
114
+ | Training Loss | Epoch | Step | Validation Loss |
115
+ |:-------------:|:-------:|:----:|:---------------:|
116
+ | No log | 0.0769 | 1 | 10.3764 |
117
+ | 10.3522 | 3.0769 | 40 | 10.3366 |
118
+ | 10.1177 | 6.1538 | 80 | 10.0885 |
119
+ | 9.8887 | 9.2308 | 120 | 9.8677 |
120
+ | 9.688 | 12.3077 | 160 | 9.6724 |
121
+ | 9.5151 | 15.3846 | 200 | 9.5050 |
122
+ | 9.3725 | 18.4615 | 240 | 9.3687 |
123
+ | 9.2678 | 21.5385 | 280 | 9.2734 |
124
+ | 9.2101 | 24.6154 | 320 | 9.2205 |
125
+ | 9.186 | 27.6923 | 360 | 9.2018 |
126
+ | 9.18 | 30.7692 | 400 | 9.1964 |
127
+ | 9.1787 | 33.8462 | 440 | 9.1945 |
128
+ | 9.1768 | 36.9231 | 480 | 9.1941 |
129
+ | 9.1775 | 40.0 | 520 | 9.1938 |
130
+ | 9.1784 | 43.0769 | 560 | 9.1949 |
131
+ | 9.1762 | 46.1538 | 600 | 9.1944 |
132
 
133
 
134
  ### Framework versions
135
 
136
  - Transformers 4.48.1
137
+ - Pytorch 2.5.1+cu124
138
  - Datasets 3.2.0
139
  - Tokenizers 0.21.0