# ─── Configuração para Axolotl fine-tuning full / SFT ───

base_model: meta-llama/Llama-3.2-3B-Instruct
tokenizer_type: AutoTokenizer
trust_remote_code: true
is_llama_derived_model: true

# dataset: lista de conjuntos que você quer usar
datasets:
  - path: nvidia/Llama-Nemotron-Post-Training-Dataset
    type:
      system_prompt: ""
      field_instruction: input
      field_input: ""
      field_output: output
      format: |-
        {instruction}
    split: chat

val_set_size: 0.01     # ou outro valor que fizer sentido
# (ou se você preferir: definir explicitamente um dataset de validação)

# Batch / treino / otimização
micro_batch_size: 1
gradient_accumulation_steps: 8
# ajuste conforme sua GPU / memória
sequence_len: 8192              # ou outro contexto máximo desejado
eval_sequence_len: 8192
pad_to_sequence_len: true        # útil se seu dataset tiver diferentes comprimentos
sample_packing: true             # útil para efficiency, dependendo do dataset


optimizer: adamw_torch_fused        # ou outro disponível
learning_rate: 2.0e-5
weight_decay: 0.0
betas: [0.9, 0.999]
eps: 1.0e-8

lr_scheduler: cosine
warmup_steps: 100

bf16: true                      # ou fp16 conforme sua infraestrutura
tf32: true
gradient_checkpointing: true

special_tokens:
  eos_token: "<|eot_id|>"
  pad_token: "<|eot_id|>"

eot_tokens:
  - "<|eot_id|>"

roles_to_train:
  - assistant

train_on_eos: last

# Salvamento / checkpoints
save_strategy: steps
save_steps: 100
save_total_limit: 3
save_safetensors: true
load_best_model_at_end: true
metric_for_best_model: eval_loss
greater_is_better: false

logging_steps: 10
output_dir: ./outputs/llama32_full_sft_instruct_data_nemotron
seed: 42

outputs/llama32_full_sft_instruct_data_nemotron

This model is a fine-tuned version of meta-llama/Llama-3.2-3B-Instruct on the nvidia/Llama-Nemotron-Post-Training-Dataset dataset.

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
gradient_accumulation_steps: 8
total_train_batch_size: 8
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: 100
training_steps: 849

Training results

Framework versions

Transformers 4.57.1
Pytorch 2.9.0+cu130
Datasets 4.3.0
Tokenizers 0.22.1

cemig-temp/llama3.2-3b-instruct-data-nemotron

outputs/llama32_full_sft_instruct_data_nemotron

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

cemig-temp/
llama3.2-3b-instruct-data-nemotron