See axolotl config

axolotl version: 0.4.1

yaml

1adapter: lora 2base_model: Qwen/Qwen2-0.5B-Instruct 3bf16: true 4chat_template: llama3 5dataset_prepared_path: null 6datasets: 7- data_files: 8 - f57d828564838a69_train_data.json 9 ds_type: json 10 format: custom 11 path: /workspace/input_data/f57d828564838a69_train_data.json 12 type: 13 field_input: chosen 14 field_instruction: source 15 field_output: reject 16 format: '{instruction} {input}' 17 no_input_format: '{instruction}' 18 system_format: '{system}' 19 system_prompt: '' 20debug: null 21device_map: 22 ? '' 23 : 0,1,2,3,4,5,6,7 24early_stopping_patience: 2 25eval_max_new_tokens: 128 26eval_steps: 100 27eval_table_size: null 28flash_attention: true 29gradient_accumulation_steps: 8 30gradient_checkpointing: true 31group_by_length: false 32hub_model_id: Alphatao/8ea1178f-7714-4a61-9d8c-478a84876cab 33hub_repo: null 34hub_strategy: null 35hub_token: null 36learning_rate: 0.0002 37load_best_model_at_end: true 38load_in_4bit: false 39load_in_8bit: false 40local_rank: null 41logging_steps: 1 42lora_alpha: 128 43lora_dropout: 0.3 44lora_fan_in_fan_out: null 45lora_model_dir: null 46lora_r: 64 47lora_target_linear: true 48lr_scheduler: cosine 49max_grad_norm: 1.0 50max_steps: 12288 51micro_batch_size: 4 52mlflow_experiment_name: /tmp/f57d828564838a69_train_data.json 53model_type: AutoModelForCausalLM 54num_epochs: 2 55optimizer: adamw_bnb_8bit 56output_dir: miner_id_24 57pad_to_sequence_len: true 58resume_from_checkpoint: null 59s2_attention: null 60sample_packing: false 61save_steps: 100 62sequence_len: 1024 63strict: false 64tf32: true 65tokenizer_type: AutoTokenizer 66train_on_inputs: false 67trust_remote_code: true 68val_set_size: 0.00648773493710141 69wandb_entity: null 70wandb_mode: online 71wandb_name: 64067547-49fe-44ed-9f15-08d6f4ccfab7 72wandb_project: Gradients-On-Demand 73wandb_run: your_name 74wandb_runid: 64067547-49fe-44ed-9f15-08d6f4ccfab7 75warmup_steps: 10 76weight_decay: 0.0 77xformers_attention: null 78

8ea1178f-7714-4a61-9d8c-478a84876cab

• Loss: 0.6288

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 0.0002
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• 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: 10
• training_steps: 12288

Training results

Framework versions

• PEFT 0.13.2
• Transformers 4.46.0
• Pytorch 2.5.0+cu124
• Datasets 3.0.1
• Tokenizers 0.20.1