---
license: apache-2.0
datasets:
- abacusai/SystemChat-1.1
language:
- en
library_name: transformers
tags:
- llama-factory
- unsloth
---
# h2o-danube2 with ChatML template

This is a [BAdam](https://arxiv.org/abs/2404.02827 "BAdam: A Memory Efficient Full Parameter Optimization Method for Large Language Models") and [LoRA+](https://arxiv.org/abs/2402.12354 "LoRA+: Efficient Low Rank Adaptation of Large Models") fine-tuned danube2 base model. It uses the ChatML template and was trained on the [SystemChat-1.1](https://huggingface.co/datasets/abacusai/SystemChat-1.1) from [Abacus.AI](https://huggingface.co/abacusai).

## Quants

Thank you [mradermacher](https://huggingface.co/mradermacher)!

- [mradermacher/danube2-1.8b-SystemChat-1.1-GGUF](https://huggingface.co/mradermacher/danube2-1.8b-SystemChat-1.1-GGUF)

## Template

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

## BAdam

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

### method
stage: sft
do_train: true
finetuning_type: full
use_badam: true
badam_switch_mode: descending
badam_switch_interval: 50
badam_start_block: 22
badam_mask_mode: scatter
badam_verbose: 1
seed: 314

### dataset
dataset: systemchat11
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: systemchat11-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: 0.00002
num_train_epochs: 3
lr_scheduler_type: cosine
warmup_ratio: 0.01
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 |
|:-------------:|:------:|:----:|:---------------:|
| 1.0062        | 0.8324 | 1000 | 0.9837          |
| 0.8484        | 1.6648 | 2000 | 0.9388          |
| 0.7834        | 2.4971 | 3000 | 0.9309          |


## QLoRA+

```yaml
### model
model_name_or_path: systemchat11-chatml-badam

### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
loraplus_lr_ratio: 16.0
lora_rank: 8
lora_alpha: 16
use_unsloth: true
quantization_bit: 4
upcast_layernorm: true
seed: 31415

### dataset
dataset: systemchat11
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: systemchat11-chatml-badam/loraplus
logging_steps: 1
save_steps: 1
save_strategy: epoch
plot_loss: true
overwrite_output_dir: false

### train
per_device_train_batch_size: 4
gradient_accumulation_steps: 4
learning_rate: 0.0001
num_train_epochs: 2.0
lr_scheduler_type: cosine
warmup_ratio: 0.01
bf16: true
flash_attn: fa2

### eval
val_size: 0.02
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
```

### QLoRA+ Training results

| Training Loss | Epoch  | Step | Validation Loss |
|:-------------:|:------:|:----:|:---------------:|
| 0.8591        | 0.4204 | 500  | 0.8457          |
| 0.9098        | 0.8409 | 1000 | 0.8251          |
| 0.735         | 1.2613 | 1500 | 0.8304          |
| 0.6811        | 1.6817 | 2000 | 0.8252          |