See axolotl config

axolotl version: 0.11.0.dev0

base_model: minpeter/tiny-ko-124m-base

hub_model_id: minpeter/tiny-ko-124m-sft
output_dir: ./outputs/tiny-ko-124m-sft
wandb_project: "axolotl"
wandb_entity: "kasfiekfs-e"

model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

strict: false

chat_template: chatml
datasets:
  - path: lemon-mint/Korean-FineTome-100k
    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: 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: 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: 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

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

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

dataset_prepared_path: last_run_prepared
val_set_size: 0.001
save_safetensors: true
sequence_len: 2048
sample_packing: false
pad_to_sequence_len: false
use_pose: true
pose_max_context_len: 65536

overrides_of_model_config:
  rope_theta: 10000.0
  max_position_embeddings: 65536

gradient_accumulation_steps: 8
micro_batch_size: 32
num_epochs: 1
optimizer: adamw_bnb_8bit
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:

tiny-ko-124m-sft

This model is a fine-tuned version of minpeter/tiny-ko-124m-base on the lemon-mint/Korean-FineTome-100k, the lemon-mint/smol-koreantalk, the heegyu/open-korean-instructions-v20231020, the trillionlabs/multisystem-curated, the allenai/tulu-3-sft-personas-instruction-following, 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: 1.7098

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: 32
• eval_batch_size: 32
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 8
• total_train_batch_size: 512
• total_eval_batch_size: 64
• optimizer: Use OptimizerNames.ADAMW_BNB 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: 5042

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0	0	2.7016
2.1419	0.0397	200	2.1320
2.0675	0.0793	400	2.0446
2.0252	0.1190	600	1.9864
1.9304	0.1587	800	1.9468
1.9536	0.1983	1000	1.9145
1.8692	0.2380	1200	1.8879
1.8556	0.2777	1400	1.8645
1.8421	0.3174	1600	1.8433
1.9118	0.3570	1800	1.8256
1.7791	0.3967	2000	1.8090
1.8162	0.4364	2200	1.7934
1.796	0.4760	2400	1.7795
1.749	0.5157	2600	1.7661
1.7536	0.5554	2800	1.7540
1.7672	0.5950	3000	1.7432
1.7523	0.6347	3200	1.7336
1.7074	0.6744	3400	1.7259
1.7218	0.7141	3600	1.7202
1.6928	0.7537	3800	1.7158
1.7184	0.7934	4000	1.7127
1.761	0.8331	4200	1.7109
1.7481	0.8727	4400	1.7101
1.7245	0.9124	4600	1.7098
1.7076	0.9521	4800	1.7097
1.7403	0.9917	5000	1.7098

Framework versions

• Transformers 4.52.4
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.1