This is a BAdam fine-tuned and LoRA+ danube2 base model. It uses the ChatML template and was trained on the Airoboros-3.2 (CC BY 4.0) dataset from jondurbin.

Quants

Thank you mradermacher!

mradermacher/danube2-1.8b-airoboros-3.2-GGUF

Template

jinja

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

BAdam

System: You are a helpful assistant.

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: 13
badam_mask_mode: scatter
seed: 314

### dataset
dataset: airoboros32
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: airoboros32-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
0.9124	0.2753	1000	0.9466
0.8072	0.5506	2000	0.9149
0.9017	0.8258	3000	0.8982
0.8883	1.1011	4000	0.8844
0.8405	1.3764	5000	0.8786
0.864	1.6517	6000	0.8754
0.7758	1.9270	7000	0.8752

QLoRA+

System: None

yaml

### model
model_name_or_path: airoboros32-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: 314

### dataset
dataset: airoboros32
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: airoboros32-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: 1000

QLoRA+ Training results

Training Loss	Epoch	Step	Validation Loss
0.9691	0.2781	1000	0.8704
0.7387	0.5562	2000	0.8443
0.6769	0.8343	3000	0.8250
0.5156	1.1123	4000	0.8134
0.4142	1.3904	5000	0.8029
0.6328	1.6685	6000	0.7953
0.872	1.9466	7000	0.7927

trollek/danube2-1.8b-airoboros-3.2

h2o-danube2 with ChatML template