File size: 2,052 Bytes
aa55a51
 
 
 
 
 
 
 
0cc5ff4
 
 
aa55a51
0cc5ff4
65a6112
0cc5ff4
65a6112
bd32b5b
 
 
 
 
 
65a6112
 
 
0cc5ff4
65a6112
0cc5ff4
 
65a6112
 
0cc5ff4
65a6112
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0cc5ff4
65a6112
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0cc5ff4
65a6112
0cc5ff4
 
 
 
 
 
 
aa55a51
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
---
license: apache-2.0
datasets:
- mlabonne/orpo-dpo-mix-40k
language:
- en
library_name: transformers
base_model: h2oai/h2o-danube2-1.8b-base
tags:
- llama-factory
- unsloth
---
# h2o-danube2 with ChatML template

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 [mlabonne/orpo-dpo-mix-40k](https://huggingface.co/datasets/mlabonne/orpo-dpo-mix-40k), but as SFT and not DPO, using LLama-Factory.

## Quants

Much love, [mradermacher](https://huggingface.co/mradermacher)!

- [mradermacher/danube2-1.8b-Neural-GGUF](https://huggingface.co/mradermacher/danube2-1.8b-Neural-GGUF)

## Template

```jinja
<|im_start|>user
{{instruction}}<|im_end|>
<|im_start|>assistant
{{response}}<|im_end|>
```

## BAdam config

```yaml
### model
model_name_or_path: danube2-base-chatml

### method
stage: sft
do_train: true
finetuning_type: full
use_badam: true
badam_switch_mode: ascending
badam_switch_interval: 50
badam_verbose: 1
badam_start_block: 12
badam_mask_mode: scatter
seed: 314

### dataset
dataset: orpo_sft_mix_40k
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: orpo-chatml-badam
logging_steps: 5
save_steps: 1
save_strategy: epoch
plot_loss: true
overwrite_output_dir: false

### train
per_device_train_batch_size: 2
gradient_accumulation_steps: 8
learning_rate: 0.00001
num_train_epochs: 2
lr_scheduler_type: cosine
warmup_ratio: 0.01
pure_bf16: true
flash_attn: fa2

### eval
val_size: 0.01
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 1000
```

### BAdam training results

| Training Loss | Epoch  | Step | Validation Loss |
|:-------------:|:------:|:----:|:---------------:|
| 0.7474        | 0.3653 | 1000 | 0.8887          |
| 0.9106        | 0.7306 | 2000 | 0.8681          |
| 0.8121        | 1.0958 | 3000 | 0.8635          |
| 0.8636        | 1.4611 | 4000 | 0.8562          |
| 0.8           | 1.8264 | 5000 | 0.8565          |