See axolotl config

axolotl version: 0.13.0.dev0

base_model: kajuma/DiffLlama-1B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

hub_model_id: 
hub_strategy: 
push_dataset_to_hub:
hf_use_auth_token: true

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

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: tokenizer_default

datasets:
  - path: kajuma/Zero_SFT_Ja_v3.5
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content

shuffle_merged_datasets: true
dataset_prepared_path: ./output/dataset
val_set_size: 0.002
output_dir: ./output/model

sequence_len: 4096
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:
lora_fan_in_fan_out:

wandb_project: diffllama
wandb_entity: tepic
wandb_watch:
wandb_name: diffllama-sft-datapilot
wandb_log_model:

gradient_accumulation_steps: 32
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch
lr_scheduler: cosine
cosine_min_lr_ratio: 0.1
learning_rate: 5e-4

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: false
early_stopping_patience:
auto_resume_from_checkpoints: true
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: false

save_strategy: steps
save_steps: 100
save_total_limit: 1

warmup_steps: 20
eval_steps: 100
eval_batch_size: 4
eval_table_size:
eval_max_new_tokens:
debug:
deepspeed:
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:

output/model

This model is a fine-tuned version of kajuma/DiffLlama-1B on the kajuma/Zero_SFT_Ja_v3.5 dataset. It achieves the following results on the evaluation set:

• Loss: 1.7823
• Ppl: 5.9437
• Memory/max Active (gib): 26.29
• Memory/max Allocated (gib): 26.29
• Memory/device Reserved (gib): 27.83

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: 0.0005
• train_batch_size: 1
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 32
• total_train_batch_size: 32
• 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: 575

Training results

Framework versions

• Transformers 4.57.1
• Pytorch 2.8.0+cu128
• Datasets 4.4.1
• Tokenizers 0.22.1