axolotl version: 0.12.0.dev0

base_model: minpeter/tiny-ko-187m-base-250725

hub_model_id: minpeter/tiny-ko-187m-sft-250725
output_dir: ./outputs/tiny-ko-187m-sft-250725
wandb_project: "axolotl"
wandb_entity: "kasfiekfs-e"

model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

strict: false

chat_template: chatml
datasets:
  - path: HuggingFaceTB/smol-smoltalk
    type: chat_template
    split: train
    field_messages: messages
    message_property_mappings:
      role: role
      content: content

  - path: trillionlabs/multisystem-curated
    type: chat_template
    split: train
    field_messages: messages
    message_property_mappings:
      role: role
      content: content

  - path: allenai/tulu-3-sft-personas-instruction-following
    type: chat_template
    split: train
    field_messages: messages
    message_property_mappings:
      role: role
      content: content

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

  - path: lemon-mint/Korean-FineTome-100k
    type: chat_template
    split: train
    field_messages: messages
    message_property_mappings:
      role: role
      content: content

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

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

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

dataset_prepared_path: last_run_prepared
val_set_size: 0.001
save_safetensors: true
sequence_len: 8192
sample_packing: false
pad_to_sequence_len: false
use_pose: true
pose_max_context_len: 65536

overrides_of_model_config:
  rope_theta: 1000000.0
  max_position_embeddings: 65536

gradient_accumulation_steps: 8
micro_batch_size: 16
num_epochs: 1
optimizer: muon
lr_scheduler: cosine
learning_rate: 3e-4

train_on_inputs: false
group_by_length: false
bf16: true
fp16:
tf32: true

gradient_checkpointing: false
gradient_checkpointing_kwargs:
  use_reentrant: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
sdp_attention:
s2_attention:

save_steps: 200
warmup_steps: 20
eval_steps: 200
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:

special_tokens:
  eos_token: '<|im_end|>'

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
  - axolotl.integrations.liger.LigerPlugin
#   - axolotl.integrations.lm_eval.LMEvalPlugin

# lm_eval_tasks:
#   - gsm8k
#   - hellaswag
#   - arc_easy
#   - arc_challenge
#   - piqa
#   - winogrande
#   - openbookqa
#   - wsc
#   - boolq

liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true

tiny-ko-187m-sft-250725

This model is a fine-tuned version of minpeter/tiny-ko-187m-base-250725 on the HuggingFaceTB/smol-smoltalk, the trillionlabs/multisystem-curated, the allenai/tulu-3-sft-personas-instruction-following, the lemon-mint/smol-koreantalk, the lemon-mint/Korean-FineTome-100k, the heegyu/open-korean-instructions-v20231020, the coastral/korean-writing-style-instruct and the devngho/korean-instruction-mix datasets. It achieves the following results on the evaluation set:

• Loss: 1.6810

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: 0.0003
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• optimizer: Use OptimizerNames.ADAMW_TORCH 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: 13880

Training results

Framework versions

• Transformers 4.53.2
• Pytorch 2.7.1+cu126
• Datasets 4.0.0
• Tokenizers 0.21.2