trollek/danube2-1.8b-Tess-v1.5

h2o-danube2 with ChatML template

This model was first fine-tuned with BAdam on migtissera/Tess-v1.5 using LLama-Factory.

Quants

Thanks to mradermacher for this!

• mradermacher/danube2-1.8b-Tess-v1.5-GGUF

Template

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

BAdam config

### 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: 6
seed: 720

### dataset
dataset: tess15
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: tess15-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: 4
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: 1000

BAdam training results

Training Loss	Epoch	Step	Validation Loss
0.8017	0.0643	1000	0.6820
0.6167	0.1287	2000	0.6610
0.6161	0.1930	3000	0.6496
0.6322	0.2574	4000	0.6423
0.5127	0.3217	5000	0.6366
0.61	0.3860	6000	0.6312
0.6758	0.4504	7000	0.6266
0.5901	0.5147	8000	0.6215
0.5163	0.5791	9000	0.6197
0.6043	0.6434	10000	0.6175
0.5056	0.7077	11000	0.6153
0.5772	0.7721	12000	0.6126
0.6692	0.8364	13000	0.6107
0.5262	0.9008	14000	0.6066
0.6386	0.9651	15000	0.6056