cemig-temp/llama3.2-3b-instruct-tokenlab-cemigConvo

See axolotl config

axolotl version: 0.13.0.dev0

base_model: meta-llama/Llama-3.2-3B-Instruct
trust_remote_code: true
strict: false

chat_template: llama3

load_in_8bit: false
load_in_4bit: false

plugins:
  - axolotl.integrations.liger.LigerPlugin

liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true

datasets:
  - path: ./outputs/dataset_tokenlab/train
    type: chat_template
    weight: 0.8
  - path: ./outputs/dataset_cemig/train
    type: chat_template
    weight: 0.2

validation_datasets:
  - path: ./outputs/dataset_tokenlab/validation
    type: chat_template
    weight: 0.8
  - path: ./outputs/dataset_cemig/validation
    type: chat_template
    weight: 0.2

test_datasets:
  - path: ./outputs/dataset_tokenlab/test
    type: chat_template
    weight: 0.8
  - path: ./outputs/dataset_cemig/test
    type: chat_template
    weight: 0.2

val_set_size: 0.0

dataset_prepared_path: ./outputs/dataset_prepared
output_dir: ./outputs/cemig-sft-2gpu/

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true
flash_attn: true

gradient_checkpointing: true
micro_batch_size: 4
gradient_accumulation_steps: 4

num_epochs: 2
optimizer: adamw_torch_fused
learning_rate: 1.0e-5
lr_scheduler: cosine
cosine_constant_lr_ratio: 0
cosine_min_lr_ratio: 0.1
warmup_ratio: 0.1
weight_decay: 0.0

bf16: true
tf32: true
save_only_model: true

logging_steps: 1
evals_per_epoch: 4
saves_per_epoch: 2

special_tokens:
  pad_token: <|finetune_right_pad_id|>

dataloader_num_workers: 4
dataloader_prefetch_factor: 2

wandb_project: llama32-3b-dados-cemig
wandb_entity: null
wandb_name: llama3.2-3b-tokenlab-more-cemig-data
wandb_log_model: checkpoint

outputs/cemig-sft-2gpu/

This model is a fine-tuned version of meta-llama/Llama-3.2-3B-Instruct on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.7199
• Memory/max Active (gib): 47.38
• Memory/max Allocated (gib): 47.38
• Memory/device Reserved (gib): 47.9

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: 1e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• total_eval_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: 3136
• training_steps: 31360

Training results

Training Loss	Epoch	Step	Validation Loss	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	2.6370	33.94	33.94	34.49
0.7927	0.2500	3920	0.8321	47.38	47.38	49.42
0.7541	0.5000	7840	0.7620	47.38	47.38	47.9
0.7028	0.7500	11760	0.7397	47.38	47.38	47.9
0.6949	1.0	15680	0.7298	47.38	47.38	47.9
0.6793	1.2500	19600	0.7247	47.38	47.38	47.9
0.71	1.5000	23520	0.7221	47.38	47.38	47.9
0.6818	1.7500	27440	0.7207	47.38	47.38	47.9
0.6828	2.0	31360	0.7199	47.38	47.38	47.9

Framework versions

• Transformers 4.57.1
• Pytorch 2.9.0+cu130
• Datasets 4.3.0
• Tokenizers 0.22.1