Commit
•
073e91a
1
Parent(s):
4e2cd1f
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: llama3
|
3 |
+
base_model: Magpie-Align/Llama-3-8B-Ultrachat-200K
|
4 |
+
tags:
|
5 |
+
- axolotl
|
6 |
+
- generated_from_trainer
|
7 |
+
model-index:
|
8 |
+
- name: Llama-3-8B-Ultrachat-200K
|
9 |
+
results: []
|
10 |
+
pipeline_tag: text-generation
|
11 |
+
---
|
12 |
+
|
13 |
+
# QuantFactory/Llama-3-8B-Ultrachat-200K-GGUF
|
14 |
+
This is quantized version of [Magpie-Align/Llama-3-8B-Ultrachat-200K](https://huggingface.co/Magpie-Align/Llama-3-8B-Ultrachat-200K) created using llama.cpp
|
15 |
+
|
16 |
+
# Model Description
|
17 |
+
|
18 |
+
[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
|
19 |
+
<details><summary>See axolotl config</summary>
|
20 |
+
|
21 |
+
axolotl version: `0.4.0`
|
22 |
+
```yaml
|
23 |
+
base_model: meta-llama/Meta-Llama-3-8B
|
24 |
+
model_type: LlamaForCausalLM
|
25 |
+
tokenizer_type: AutoTokenizer
|
26 |
+
|
27 |
+
load_in_8bit: false
|
28 |
+
load_in_4bit: false
|
29 |
+
strict: false
|
30 |
+
|
31 |
+
datasets:
|
32 |
+
- path: PhilipMay/UltraChat-200k-ShareGPT-clean
|
33 |
+
type: sharegpt
|
34 |
+
conversation: llama3
|
35 |
+
dataset_prepared_path: last_run_prepared
|
36 |
+
val_set_size: 0.001
|
37 |
+
output_dir: ./out_Llama-8B-Ultrachat-200K
|
38 |
+
|
39 |
+
sequence_len: 8192
|
40 |
+
sample_packing: true
|
41 |
+
eval_sample_packing: false
|
42 |
+
pad_to_sequence_len: true
|
43 |
+
|
44 |
+
wandb_project: SynDa
|
45 |
+
wandb_entity:
|
46 |
+
wandb_watch:
|
47 |
+
wandb_name: Llama-3-8B-Ultrachat-200K
|
48 |
+
wandb_log_model:
|
49 |
+
hub_model_id: SynDa/Llama-3-8B-Ultrachat-200K
|
50 |
+
|
51 |
+
gradient_accumulation_steps: 8
|
52 |
+
micro_batch_size: 1
|
53 |
+
num_epochs: 2
|
54 |
+
optimizer: paged_adamw_8bit
|
55 |
+
lr_scheduler: cosine
|
56 |
+
learning_rate: 2e-5
|
57 |
+
|
58 |
+
train_on_inputs: false
|
59 |
+
group_by_length: false
|
60 |
+
bf16: auto
|
61 |
+
fp16:
|
62 |
+
tf32: false
|
63 |
+
|
64 |
+
gradient_checkpointing: true
|
65 |
+
gradient_checkpointing_kwargs:
|
66 |
+
use_reentrant: false
|
67 |
+
early_stopping_patience:
|
68 |
+
resume_from_checkpoint:
|
69 |
+
logging_steps: 1
|
70 |
+
xformers_attention:
|
71 |
+
flash_attention: true
|
72 |
+
|
73 |
+
warmup_steps: 100
|
74 |
+
evals_per_epoch: 1
|
75 |
+
eval_table_size:
|
76 |
+
saves_per_epoch: 3
|
77 |
+
debug:
|
78 |
+
deepspeed:
|
79 |
+
weight_decay: 0.0
|
80 |
+
fsdp:
|
81 |
+
fsdp_config:
|
82 |
+
special_tokens:
|
83 |
+
pad_token: <|end_of_text|>
|
84 |
+
|
85 |
+
```
|
86 |
+
|
87 |
+
</details><br>
|
88 |
+
|
89 |
+
# Llama-3-8B-Ultrachat-200K
|
90 |
+
|
91 |
+
This model is a fine-tuned version of [meta-llama/Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B) on the None dataset.
|
92 |
+
It achieves the following results on the evaluation set:
|
93 |
+
- Loss: 0.8219
|
94 |
+
|
95 |
+
## Training procedure
|
96 |
+
|
97 |
+
### Training hyperparameters
|
98 |
+
|
99 |
+
The following hyperparameters were used during training:
|
100 |
+
- learning_rate: 2e-05
|
101 |
+
- train_batch_size: 1
|
102 |
+
- eval_batch_size: 1
|
103 |
+
- seed: 42
|
104 |
+
- distributed_type: multi-GPU
|
105 |
+
- num_devices: 4
|
106 |
+
- gradient_accumulation_steps: 8
|
107 |
+
- total_train_batch_size: 32
|
108 |
+
- total_eval_batch_size: 4
|
109 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
110 |
+
- lr_scheduler_type: cosine
|
111 |
+
- lr_scheduler_warmup_steps: 100
|
112 |
+
- num_epochs: 2
|
113 |
+
|
114 |
+
### Training results
|
115 |
+
|
116 |
+
| Training Loss | Epoch | Step | Validation Loss |
|
117 |
+
|:-------------:|:------:|:----:|:---------------:|
|
118 |
+
| 1.0564 | 0.0010 | 1 | 1.0959 |
|
119 |
+
| 0.857 | 0.9995 | 1016 | 0.8206 |
|
120 |
+
| 0.7924 | 1.9785 | 2032 | 0.8219 |
|
121 |
+
|
122 |
+
|
123 |
+
### Framework versions
|
124 |
+
|
125 |
+
- Transformers 4.40.2
|
126 |
+
- Pytorch 2.3.0+cu121
|
127 |
+
- Datasets 2.19.1
|
128 |
+
- Tokenizers 0.19.1
|