See axolotl config

axolotl version: 0.8.0.dev0

base_model: halcyon-llm/Llama-halcyon-1B-base-checkpoint-20480
tokenizer_config: meta-llama/Llama-3.2-1B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# 学習後のモデルのHFへのアップロードに関する設定
hub_model_id: kajuma/Llama-halcyon-1B-base-sft
hub_strategy: "end"
push_dataset_to_hub:
hf_use_auth_token: true

# 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

load_in_8bit: false
load_in_4bit: true

chat_template: llama3

datasets:
  - path: Aratako/Magpie-Tanuki-Qwen2.5-72B-Answered
    split: train
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
  - path: kanhatakeyama/ramdom-to-fixed-multiturn-Calm3
    split: 20240806filtered
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/qlora-out

adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

sequence_len: 2048
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

wandb_project: halcyon-instruct
wandb_entity: tepic
wandb_watch:
wandb_name: sft-lora-1
wandb_log_model:

gradient_accumulation_steps: 16
micro_batch_size: 1
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
cosine_min_lr_ratio: 0.1
learning_rate: 0.0002

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

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

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 70
eval_steps: 300
eval_batch_size: 1
eval_table_size:
eval_max_new_tokens:
debug:
weight_decay: 0.01
special_tokens:
  pad_token: "<|end_of_text|>"

Llama-halcyon-1B-base-sft

This model is a fine-tuned version of halcyon-llm/Llama-halcyon-1B-base-checkpoint-20480 on the Aratako/Magpie-Tanuki-Qwen2.5-72B-Answered and the kanhatakeyama/ramdom-to-fixed-multiturn-Calm3 datasets. It achieves the following results on the evaluation set:

• Loss: 0.9778

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: 0.0002
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 16
• 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: 70
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.5799	0.0006	1	1.6489
1.0229	0.1680	300	1.0472
1.0495	0.3360	600	1.0141
1.0596	0.5039	900	0.9952
1.0084	0.6719	1200	0.9834
1.0163	0.8399	1500	0.9778

Framework versions

• PEFT 0.15.0
• Transformers 4.50.0
• Pytorch 2.6.0+cu124
• Datasets 3.4.1
• Tokenizers 0.21.1