trollek/danube2-1.8b-MathInstruct

h2o-danube2 with ChatML template

This model was first fine-tuned with BAdam on TIGER-Lab/MathInstruct using LLama-Factory.

Quants

Mad props, mradermacher!

• mradermacher/danube2-1.8b-MathInstruct-GGUF

Template

<|im_start|>system
You are a helpful assistant specialised in mathematics.<|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: 7
seed: 5772

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

### output
output_dir: mathinstruct-chatml-badam
logging_steps: 5
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: 0.000005
num_train_epochs: 1
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
0.2748	0.0617	1000	0.2788
0.2786	0.1234	2000	0.2503
0.18	0.1850	3000	0.2144
0.2015	0.2467	4000	0.1926
0.2044	0.3084	5000	0.1777
0.142	0.3701	6000	0.1661
0.1813	0.4317	7000	0.1570
0.1413	0.4934	8000	0.1529
0.1805	0.5551	9000	0.1462
0.1431	0.6168	10000	0.1410
0.1693	0.6784	11000	0.1375
0.1291	0.7401	12000	0.1357
0.1501	0.8018	13000	0.1348
0.1521	0.8635	14000	0.1345
0.1279	0.9251	15000	0.1346
0.1351	0.9868	16000	0.1344

GSM8K results

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
gsm8k	3	strict-match	5	exact_match	0.2691	±	0.0122
		flexible-extract	5	exact_match	0.2752	±	0.0123

It matches the chat trained model from h2o.