trollek/danube2-1.8b-Code-290k

h2o-danube2 with ChatML template

This model was first fine-tuned with BAdam on ajibawa-2023/Code-290k-ShareGPT using LLama-Factory.

Template

<|im_start|>system
You are a helpful coding assistant.<|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: 8
seed: 8

### dataset
dataset: code_290k
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: code-290k-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: 1000

BAdam training results

Training Loss	Epoch	Step	Validation Loss
0.7404	0.0559	1000	0.7784
0.7858	0.1118	2000	0.7702
0.7274	0.1677	3000	0.7604
0.6956	0.2236	4000	0.7570
0.7711	0.2795	5000	0.7541
0.7643	0.3354	6000	0.7518
0.8255	0.3913	7000	0.7496
0.7456	0.4472	8000	0.7483
0.7718	0.5031	9000	0.7447
0.6693	0.5590	10000	0.7445
0.7409	0.6149	11000	0.7433
0.7319	0.6709	12000	0.7424
0.7636	0.7268	13000	0.7415
0.7504	0.7827	14000	0.7414
0.7735	0.8386	15000	0.7374
0.7438	0.8945	16000	0.7375
0.839	0.9504	17000	0.7373