giux78/zagreus-test-202000-sft-3

See axolotl config

axolotl version: 0.9.2

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

#tokenizer_type: auto
strict: false
seed: 1337
output_dir: ./ale_outputs/zagreus-350M-sft

# === Datasets ===
streaming: false
#datasets:
#  - path: /leonardo_work/EUHPC_A04_045/training/sft_data #/leonardo_work/EUHPC_A04_045/training/test_data #/leonardo_work/EUHPC_A04_045/.data
#    type: chat_template
#    chat_template: tokenizer_default_fallback_chatml
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#    roles:
#      user: ["human","user"]
#      assistant: ["gpt","assistant"]
#      system: ["system"]
#      tool: ["tool"]
#   roles_to_train: ["assistant"]
#   train_on_eos: "turn"

default_system_message: "Sei un assistente utile."
chat_template_jinja: |
      {% if messages[0]['role'] == 'system' %}
        {% set loop_messages = messages[1:] %}
        {% set system_message = messages[0]['content'] %}
      {% else %}
        {% set loop_messages = messages %}
        {% set system_message = false %}
      {% endif %}
      {{ bos_token }}
      {% if system_message != false %}
        {{ '<|start_header_id|>' + 'system' + '<|end_header_id|>\n\n' + system_message + '<|eot_id|>' }}
      {% endif %}
      {% for message in loop_messages %}
        {% set content = message['content'] | trim %}
        {% if message['role'] == 'user' %}
          {{ '<|start_header_id|>' + 'user' + '<|end_header_id|>\n\n' + content + '<|eot_id|>' }}
        {% elif message['role'] == 'assistant' %}
          {{ '<|start_header_id|>' + 'assistant' + '<|end_header_id|>\n\n' + content + '<|eot_id|>' }}
        {% endif %}
      {% endfor %}

datasets:
  - path: /leonardo_work/EUHPC_A04_045/training/test_data_cleaned   #/leonardo_work/EUHPC_A04_045/training/sft_data
    # 'type: chat_template' is the correct type for this task.
    type: chat_template
    # Your original mapping was correct for reading your dataset format. We keep it.
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
    roles:
      user: ["human", "user"]
      assistant: ["gpt", "assistant"]
      system: ["system"]
      tool: ["tool"]
    roles_to_train: ["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

optimizer: adamw_torch_fused
#learning_rate: 5e-4              # ~scaling lineare per eff. batch 512
#lr_scheduler: cosine
#warmup_ratio: 0.01
#weight_decay: 0.10
learning_rate: 2e-5          # VALORE FONDAMENTALE: Ridotto da 5e-4 a un valore sicuro per SFT.
weight_decay: 0.01           # Regolarizzazione più leggera, standard per SFT.
warmup_ratio: 0.03           # Warmup più lungo per una maggiore stabilità iniziale.
adam_beta1: 0.9
adam_beta2: 0.95
adam_epsilon: 1e-8
max_grad_norm: 1.0


# === Batch (per GPU) ===
micro_batch_size: 1
gradient_accumulation_steps: 16 #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: 100          # 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:
  bos_token: <|begin_of_text|>
  pad_token: <|end_of_text|>
  #eos_token: <|end_of_text|>
  eos_token: <|eot_id|>
  unk_token: <unk>

tokens:
  - <|start_header_id|>
  - <|end_header_id|>
  - <|eot_id|>

ale_outputs/zagreus-350M-sft

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: 2.2716

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: 1337
• distributed_type: multi-GPU
• num_devices: 32
• gradient_accumulation_steps: 16
• total_train_batch_size: 512
• total_eval_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 2
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0283	1	2.2722
No log	0.2265	8	2.2719
No log	0.4531	16	2.2718
3.8829	0.6796	24	2.2716
3.8829	0.9062	32	2.2716

Framework versions

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