Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,95 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: h2oai/h2o-danube2-1.8b-base
|
| 4 |
+
datasets:
|
| 5 |
+
- migtissera/Tess-v1.5
|
| 6 |
+
language:
|
| 7 |
+
- en
|
| 8 |
+
library_name: transformers
|
| 9 |
+
tags:
|
| 10 |
+
- llama-factory
|
| 11 |
+
- unsloth
|
| 12 |
+
---
|
| 13 |
+
# h2o-danube2 with ChatML template
|
| 14 |
+
|
| 15 |
+
This model was first fine-tuned with [BAdam](https://arxiv.org/abs/2404.02827 "BAdam: A Memory Efficient Full Parameter Optimization Method for Large Language Models") on [migtissera/Tess-v1.5](https://huggingface.co/datasets/migtissera/Tess-v1.5) using LLama-Factory.
|
| 16 |
+
|
| 17 |
+
## Template
|
| 18 |
+
|
| 19 |
+
```jinja
|
| 20 |
+
<|im_start|>system
|
| 21 |
+
{{system}}<|im_end|>
|
| 22 |
+
<|im_start|>user
|
| 23 |
+
{{instruction}}<|im_end|>
|
| 24 |
+
<|im_start|>assistant
|
| 25 |
+
{{response}}<|im_end|>
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
## BAdam config
|
| 29 |
+
|
| 30 |
+
```yaml
|
| 31 |
+
### model
|
| 32 |
+
model_name_or_path: danube2-base-chatml
|
| 33 |
+
|
| 34 |
+
### method
|
| 35 |
+
stage: sft
|
| 36 |
+
do_train: true
|
| 37 |
+
finetuning_type: full
|
| 38 |
+
use_badam: true
|
| 39 |
+
badam_switch_mode: ascending
|
| 40 |
+
badam_switch_interval: 50
|
| 41 |
+
badam_verbose: 1
|
| 42 |
+
badam_start_block: 6
|
| 43 |
+
seed: 720
|
| 44 |
+
|
| 45 |
+
### dataset
|
| 46 |
+
dataset: tess15
|
| 47 |
+
template: hermes_chatml
|
| 48 |
+
cutoff_len: 8192
|
| 49 |
+
overwrite_cache: false
|
| 50 |
+
preprocessing_num_workers: 12
|
| 51 |
+
|
| 52 |
+
### output
|
| 53 |
+
output_dir: tess15-chatml-badam
|
| 54 |
+
logging_steps: 5
|
| 55 |
+
save_steps: 1
|
| 56 |
+
save_strategy: epoch
|
| 57 |
+
plot_loss: true
|
| 58 |
+
overwrite_output_dir: false
|
| 59 |
+
|
| 60 |
+
### train
|
| 61 |
+
per_device_train_batch_size: 2
|
| 62 |
+
gradient_accumulation_steps: 4
|
| 63 |
+
learning_rate: 0.00001
|
| 64 |
+
num_train_epochs: 1
|
| 65 |
+
lr_scheduler_type: constant_with_warmup
|
| 66 |
+
warmup_ratio: 0.01
|
| 67 |
+
bf16: true
|
| 68 |
+
flash_attn: fa2
|
| 69 |
+
|
| 70 |
+
### eval
|
| 71 |
+
val_size: 0.01
|
| 72 |
+
per_device_eval_batch_size: 1
|
| 73 |
+
eval_strategy: steps
|
| 74 |
+
eval_steps: 1000
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
### BAdam training results
|
| 78 |
+
|
| 79 |
+
| Training Loss | Epoch | Step | Validation Loss |
|
| 80 |
+
|:-------------:|:------:|:-----:|:---------------:|
|
| 81 |
+
| 0.8017 | 0.0643 | 1000 | 0.6820 |
|
| 82 |
+
| 0.6167 | 0.1287 | 2000 | 0.6610 |
|
| 83 |
+
| 0.6161 | 0.1930 | 3000 | 0.6496 |
|
| 84 |
+
| 0.6322 | 0.2574 | 4000 | 0.6423 |
|
| 85 |
+
| 0.5127 | 0.3217 | 5000 | 0.6366 |
|
| 86 |
+
| 0.61 | 0.3860 | 6000 | 0.6312 |
|
| 87 |
+
| 0.6758 | 0.4504 | 7000 | 0.6266 |
|
| 88 |
+
| 0.5901 | 0.5147 | 8000 | 0.6215 |
|
| 89 |
+
| 0.5163 | 0.5791 | 9000 | 0.6197 |
|
| 90 |
+
| 0.6043 | 0.6434 | 10000 | 0.6175 |
|
| 91 |
+
| 0.5056 | 0.7077 | 11000 | 0.6153 |
|
| 92 |
+
| 0.5772 | 0.7721 | 12000 | 0.6126 |
|
| 93 |
+
| 0.6692 | 0.8364 | 13000 | 0.6107 |
|
| 94 |
+
| 0.5262 | 0.9008 | 14000 | 0.6066 |
|
| 95 |
+
| 0.6386 | 0.9651 | 15000 | 0.6056 |
|