See axolotl config

axolotl version: 0.10.0


# base model
base_model: sbintuitions/sarashina2.2-3b-instruct-v0.1
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# model upload 
hub_model_id: OsakanaTeishoku/sarashina2.2-3b-cot-sft-step1000-test-20251006
hub_strategy: "end"
push_dataset_to_hub:
hf_use_auth_token: true

# liger kernel
plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_cross_entropy: false
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

# quantize
load_in_8bit: false
load_in_4bit: false

# chat_template 
chat_template: tokenizer_default

# Liger Kernelの設定（学習の軽量・高速化）
plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_cross_entropy: false
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

# dataset
datasets:
  - path: OsakanaTeishoku/Zero_SFT_Ja_v3_Reasoning_formatted
    split: train
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
        

dataset_prepared_path: ./dataset
#val_set_size: 0.05
output_dir: ./outputs
# Training configuration
train_on_inputs: false
group_by_length: false

sequence_len: 32768
sample_packing: false

adapter: lora
lora_r: 16
lora_alpha: 16
lora_dropout: 0
# lora_target_modules:
#   - q_proj
#   - v_proj
#   - k_proj
#   - o_proj
#   - gate_proj
#   - down_proj
#   - up_proj
lora_target_linear: true
lora_modules_to_save: [embed_tokens, lm_head] 

gradient_accumulation_steps: 4
micro_batch_size: 2
#num_epochs: 1
max_steps: 1000
optimizer: adamw_torch_fused
lr_scheduler: linear
learning_rate: 1e-4

bf16: auto
tf32: false

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

warmup_ratio: 0.1
#evals_per_epoch: 4
#saves_per_epoch: 1
save_steps: 100
save_strategy:
weight_decay: 0.0




#special_tokens:
#  pad_token: <|end_of_text|>
tokens: ["<think>", "</think>"]
overrides_of_model_config: {"rope_scaling": {"rope_type": "yarn", "factor": 4.0, "original_max_position_embeddings": 8192}}
# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

sarashina2.2-3b-cot-sft-step1000-test-20251006

This model is a fine-tuned version of sbintuitions/sarashina2.2-3b-instruct-v0.1 on the OsakanaTeishoku/Zero_SFT_Ja_v3_Reasoning_formatted dataset.

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: 0.0001
train_batch_size: 2
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 8
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 100
training_steps: 1000