See axolotl config

axolotl version: 0.10.0.dev0

base_model: minpeter/tiny-ko-base

hub_model_id: minpeter/tiny-ko-sft
output_dir: ./outputs/tiny-ko-sft
wandb_project: "axolotl"
wandb_entity: "kasfiekfs-e"

chat_template: chatml
datasets:
  - path: lemon-mint/Korean-FineTome-100k
    type: chat_template
    split: train[:10%]
    field_messages: messages
    message_property_mappings:
      role: role
      content: content

  - path: lemon-mint/smol-koreantalk
    type: chat_template
    split: train[:10%]
    field_messages: messages
    message_property_mappings:
      role: role
      content: content

  - path: heegyu/open-korean-instructions-v20231020
    type: chat_template
    split: train[:10%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
    roles:
      user: ["human", "user"]
      assistant: ["gpt", "assistant", "bot"]
      system: ["system", "input"]

  # NOTE: https://github.com/FreedomIntelligence/MultilingualSIFT
  - path: FreedomIntelligence/evol-instruct-korean
    type: chat_template
    split: train[:10%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value

  - path: FreedomIntelligence/alpaca-gpt4-korean
    type: chat_template
    split: train[:10%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value

  - path: FreedomIntelligence/sharegpt-korean
    type: chat_template
    split: train[:10%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value

  - path: coastral/korean-writing-style-instruct
    type: chat_template
    split: train[:10%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value

  - path: devngho/korean-instruction-mix
    type: chat_template
    split: train[:10%]
    field_messages: messages
    message_property_mappings:
      role: from
      content: value

  - path: youjunhyeok/Magpie-Pro-300K-Filtered-ko
    type: chat_template
    split: train[:10%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value

  - path: youjunhyeok/smoltalk-ko-translate
    type: chat_template
    name: merge_filtered
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content

dataset_prepared_path: last_run_prepared
val_set_size: 0.05

save_steps: 200
warmup_steps: 20
eval_steps: 200

sequence_len: 4096

# <<<< experimental settings <<<<
sample_packing: true
train_on_inputs: false
# >>>> experimental settings >>>

pad_to_sequence_len: true

gradient_accumulation_steps: 4
micro_batch_size: 32

optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 1e-3

bf16: auto
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

num_epochs: 1
weight_decay: 0.0

tiny-ko-sft

This model is a fine-tuned version of minpeter/tiny-ko-base on the lemon-mint/Korean-FineTome-100k, the lemon-mint/smol-koreantalk, the heegyu/open-korean-instructions-v20231020, the FreedomIntelligence/evol-instruct-korean, the FreedomIntelligence/alpaca-gpt4-korean, the FreedomIntelligence/sharegpt-korean, the coastral/korean-writing-style-instruct, the devngho/korean-instruction-mix, the youjunhyeok/Magpie-Pro-300K-Filtered-ko and the youjunhyeok/smoltalk-ko-translate datasets. It achieves the following results on the evaluation set:

• Loss: 1.5297

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.001
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 256
• total_eval_batch_size: 64
• optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 20
• training_steps: 817

Training results

Training Loss	Epoch	Step	Validation Loss
2.3518	0.0012	1	2.3640
1.6322	0.2446	200	1.6913
1.5903	0.4891	400	1.6003
1.5146	0.7337	600	1.5392
1.5277	0.9783	800	1.5297

Framework versions

• Transformers 4.52.3
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.1