trollek/danube2-1.8b-openhermes

h2o-danube2 with ChatML template

This is a BAdam and LoRA+ fine-tuned danube2 base model. It uses the ChatML template and was trained on the openhermes-unfiltered.

Quants

Thank you mradermacher!

• mradermacher/danube2-1.8b-openhermes-GGUF

Template

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

BAdam

System: You are a helpful assistant.

### 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: 10
seed: 720

### dataset
dataset: openhermes_unfiltered
template: ninja_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: openhermes-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: 1
lr_scheduler_type: constant_with_warmup
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: 2000

BAdam Training results

Training Loss	Epoch	Step	Validation Loss
0.7971	0.1748	2000	0.7418
0.6815	0.3496	4000	0.7178
0.6593	0.5245	6000	0.7055
0.6923	0.6993	8000	0.6960
0.6942	0.8741	10000	0.6877

QLoRA+

### model
model_name_or_path: openhermes-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: 3141

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

### output
output_dir: openhermes-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: 1.0
lr_scheduler_type: cosine
warmup_ratio: 0.01
bf16: true
flash_attn: fa2
#neftune_noise_alpha: 5

### eval
val_size: 0.02
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 1000

QLoRA+ Training results

Training Loss	Epoch	Step	Validation Loss
0.6523	0.0883	1000	0.7126
0.6398	0.1766	2000	0.7086
0.6865	0.2649	3000	0.7001
0.6714	0.3532	4000	0.6917
0.7213	0.4415	5000	0.6819
0.7764	0.5298	6000	0.6721
0.6931	0.6181	7000	0.6638
0.6632	0.7064	8000	0.6560
0.5966	0.7947	9000	0.6514
0.6339	0.8830	10000	0.6482
0.4987	0.9713	11000	0.6472