README.md

metadata

license: apache-2.0
datasets:
  - hiyouga/glaive-function-calling-v2-sharegpt
language:
  - en
library_name: transformers
tags:
  - llama-factory
  - unsloth
base_model: h2oai/h2o-danube2-1.8b-base

h2o-danube2 with ChatML template

This is a BAdam fine-tuned danube2 base model. It uses the ChatML template and was trained on the glaive-function-calling-v2 dataset from GlaiveAI that has been converted to ShareGPT by hiyouga of LLama-Factory fame.

Template

ChatML

<|im_start|>system
{{system}}
<tools>
{{json_format_tools}}
</tools><|im_end|>
<|im_start|>user
{{instruction}}<|im_end|>
<|im_start|>assistant
<tool_call>
{{tool_call}}
</tool_call><|im_end|>
<|im_start|>tool
<tool_response>
{{response}}
</tool_response><|im_end|>

LLama-Factory

_register_template(
    name="hermes_chatml",
    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
    format_assistant=StringFormatter(slots=["{{content}}<|im_end|>\n"]),
    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
    format_function=FunctionFormatter(slots=["<tool_call>\n{\"name\":\"{{name}}\", \"arguments\":{{arguments}}}\n</tool_call><|im_end|>\n"]),
    format_observation=StringFormatter(slots=["<|im_start|>tool\n<tool_response>\n{{content}}\n</tool_response><|im_end|>\n<|im_start|>assistant\n"]),
    format_tools=ToolFormatter(tool_format="chatml"),
    stop_words=["<|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: 5
seed: 404

### dataset
dataset: glaive_toolcall_100k
template: hermes_chatml
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: glaive-tool-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: 0.000005
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
0.3914	0.1607	1000	0.2984
0.3256	0.3214	2000	0.2819
0.4131	0.4821	3000	0.2765
0.3922	0.6428	4000	0.2736
0.3528	0.8036	5000	0.2724
0.3477	0.9643	6000	0.2724