See axolotl config

axolotl version: 0.9.2

yaml

1 # ============= SFT DEBUG (~1M conv) ============= 2base_model: giux78/zagreus-test-202000 3strict: false 4output_dir: ./ale_outputs/opendata-sft-debug 5seed: 42 6 7datasets: 8 - path: /leonardo_work/EUHPC_A04_045/training/opendata-1000000 9 type: chat_template 10 field_messages: conversation 11 roles_to_train: ["assistant"] 12 train_on_eos: turn 13 14dataset_prepared_path: ./ale_outputs/dataset_cache/opendata-sft 15#default_system_message: "Sei un assistente utile." 16chat_template: llama3 17 18sequence_len: 4096 19sample_packing: true 20eval_sample_packing: true 21pad_to_sequence_len: true 22 23 24 25# --- Cosine knobs (Axolotl) --- 26# 1) Mantieni LR costante al massimo per l'80% degli step 27cosine_constant_lr_ratio: 0.8 28# 2) Floor al 30% del LR max (min_lr = 4.5e-6) 29cosine_min_lr_ratio: 0.3 30# Non impostare lr_div_factor quando usi cosine_min_lr_ratio 31 32 33optimizer: adamw_torch_fused 34lr_scheduler: cosine # <-- per isolare il comportamento 35learning_rate: 1.5e-5 36warmup_ratio: 0.05 # un po’ più lungo in debug 37weight_decay: 0.01 38max_grad_norm: 1.0 39 40micro_batch_size: 1 41gradient_accumulation_steps: 8 42# Usa max_steps per “più step” indipendentemente dalla lunghezza effettiva del dataset 43max_steps: 3000 # ≈ 4x gli step attuali 44#num_epochs: null # ignora epoche quando max_steps è settato 45 46bf16: auto 47flash_attention: true 48gradient_checkpointing: true 49 50logging_steps: 10 51eval_strategy: steps 52eval_steps: 100 53save_strategy: steps 54save_steps: 500 55save_total_limit: 3 56val_set_size: 10000 57 58fsdp_config: 59 fsdp_sharding_strategy: FULL_SHARD 60 fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP 61 fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer 62 fsdp_backward_prefetch_policy: BACKWARD_PRE 63 fsdp_state_dict_type: FULL_STATE_DICT 64 65special_tokens: 66 pad_token: <|end_of_text|> 67 eos_token: <|end_of_text|> 68 69

ale_outputs/opendata-sft-debug

• Loss: 3.5482

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 1.5e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 32
• gradient_accumulation_steps: 8
• total_train_batch_size: 256
• total_eval_batch_size: 32
• 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: cosine
• lr_scheduler_warmup_steps: 24
• training_steps: 3000

Training results

Framework versions

• Transformers 4.56.2
• Pytorch 2.5.1+cu121
• Datasets 3.5.1
• Tokenizers 0.22.1