AlexHung29629/Draft1

See axolotl config

axolotl version: 0.12.0.dev0

base_model: out_khanacademy
remove_unused_columns: true
auto_resume_from_checkpoints: true
plugins:
  - axolotl.integrations.liger.LigerPlugin
  #- axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

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

unfrozen_parameters:
  - ^\S+layers\S+$
  - ^\S+norm\S+$


datasets:
  - path: AlexHung29629/nllb_processed
    split: train[:1_000_000]
    type: chat_template
    chat_template: jinja
    chat_template_jinja: "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{message['content'] + '\n'}}{% elif (message['role'] == 'user') %}{{'Source: ' + '\n' + message['content'] + '\n' + '\nTarget:\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '</s>' + '\n'}}{% endif %}{% endfor %}"
    roles_to_train: ['user', 'assistant']

#test_datasets:
#  - path: HuggingFaceTB/cosmopedia
#    name: khanacademy
#    split: train[-100:]
#    type:
#      system_prompt: ""
#      field_system: 
#      field_instruction: prompt
#      field_output: text
#      format: "User: {instruction}\n\nAssistant: "
#      no_input_format: "User: {instruction}\n\nAssistant: "
      
sample_packing_bin_size: 500
dataset_prepared_path: data_prep_nllb
output_dir: ./out_nllb
dataloader_num_workers: 1
dataloader_pin_memory: true
shuffle_merged_datasets: false

sequence_len: 8192
eval_sequence_len: 2048
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

use_tensorboard: true
use_wandb: true
# Set the name of your wandb run
wandb_name: nllb
# Your wandb project name
wandb_project: Draft_Tiny

gradient_accumulation_steps: 1

micro_batch_size: 1
num_epochs: 1
#eval_steps: 500
save_steps: 1000
save_total_limit: 1
save_only_model: false
optimizer: adamw_8bit
adam_beta1: 0.9
adam_beta2: 0.95
adam_epsilon: 1e-6
lr_scheduler: constant_with_warmup
learning_rate: 0.0003
max_grad_norm: 1.0

bf16: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false

torch_compile: true
torch_compile_backend: inductor
torch_compile_mode: default
#flash_attention: true
#sdp_attention: true
#xformers_attention: true
flex_attention: true
flex_attn_compile_kwargs:
  dynamic: false
  mode: max-autotune-no-cudagraphs

warmup_steps: 1
logging_steps: 1
weight_decay: 0.001


special_tokens:
  bos_token: <s>
  eos_token: </s>
  pad_token: <pad>
  unk_token: <unk>

out_nllb

This model was trained from scratch on the AlexHung29629/nllb_processed 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: 0.0003
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_8BIT with betas=(0.9,0.95) and epsilon=1e-06 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 2
• training_steps: 13786

Training results

Framework versions

• Transformers 4.54.1
• Pytorch 2.7.1+cu128
• Datasets 4.0.0
• Tokenizers 0.21.4