base_model: giux78/zagreus-test-202000
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

strict: false


# === Datasets ===
streaming: true
datasets:
  - path: /leonardo_work/EUHPC_A04_045/.data
    type: chat_template
    chat_template: tokenizer_default_fallback_llama3
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
    roles_to_train: ["gpt", "assistant"]
    train_on_eos: "turn"

# === Sequencing / packing ===
sequence_len: 4096
sample_packing: true
remove_unused_columns: false   # <-- aggiungi questa riga
eval_sample_packing: false
pad_to_sequence_len: false
streaming_multipack_buffer_size: 10000

# === Ottimizzazione ===
optimizer: adamw_torch_fused
learning_rate: 2e-5
lr_scheduler: cosine
warmup_ratio: 0.1
weight_decay: 0.0

# === Batch (per GPU) ===
micro_batch_size: 1
gradient_accumulation_steps: 8
# Eff. batch = micro_batch_size * grad_accum * num_gpus = 1 * 8 * 32 = 256

# === Precisione / memoria ===
bf16: auto
tf32: true
flash_attention: true
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false

# === FSDP (Axolotl usa fsdp_config; la chiave "fsdp:" è deprecata) ===
fsdp_config:
  fsdp_sharding_strategy: FULL_SHARD           # shard di param, grad e optimizer state
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
  fsdp_use_orig_params: false
  fsdp_sync_module_states: true
  fsdp_limit_all_gathers: true
  fsdp_cpu_ram_efficient_loading: true
  fsdp_offload_params: false                   # attivalo solo se VRAM è stretta (vedi variante)
  fsdp_state_dict_type: SHARDED_STATE_DICT     # checkpoint più leggeri su cluster multi-nodo

# === Loop di training ===
num_epochs: 1                                  # con 170GB basta una passata
# max_steps: 200000                            # alternativa: budget a step/token

# === Eval / checkpoint ===
val_set_size: 0.01
evals_per_epoch: 5
save_steps: 2000          # salva ogni 2.000 step (metti il valore che preferisci)
save_total_limit: 5
logging_steps: 20

# === Tracciamento ===
wandb_mode: "offline"
wandb_project: zagreus-350M-sft
wandb_entity: mii-llm
wandb_name: sft

# === Token speciali ===
special_tokens:
  pad_token: <|end_of_text|>
  eos_token: <|end_of_text|>

model-out

This model is a fine-tuned version of giux78/zagreus-test-202000 on the None dataset. It achieves the following results on the evaluation set:

Loss: 2.2814

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: 2e-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: 340
num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0003	1	2.1506
3.6166	0.2003	681	2.1390
3.4709	0.4006	1362	2.2716
3.4327	0.6008	2043	2.2737
3.4102	0.8011	2724	2.2814

Framework versions

Transformers 4.56.2
Pytorch 2.5.1+cu121
Datasets 3.5.1
Tokenizers 0.22.1

giux78/zagreus-test-202000-sft

model-out

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

giux78/
zagreus-test-202000-sft