README.md

metadata

library_name: transformers
tags:
  - axolotl
  - generated_from_trainer
datasets:
  - train_v2.1.jsonl
model-index:
  - name: llama33_fix3
    results: []

See axolotl config

axolotl version: 0.8.1

# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
# This can also be a relative path to a model on disk
base_model: ./step1-embed-model/merged/

# Use CUDA bf16
bf16: auto
tf32: false

# List[str]. Add plugins to extend the pipeline.
# See `src/axolotl/integrations` for the available plugins or doc below for more details.
# https://axolotl-ai-cloud.github.io/axolotl/docs/custom_integrations.html
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

# A list of one or more datasets to finetune the model with
datasets:
  # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
  - path: train_v2.1.jsonl
    # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]
    type: input_output # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>

# If false, the datasets will not be shuffled and will keep their original order in `datasets`.
# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.
shuffle_merged_datasets: false

# Axolotl attempts to save the dataset as an arrow after packing the data together so
# subsequent training attempts load faster, relative path
dataset_prepared_path: data/llama33fix_prepared
# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
# if not set.
dataset_processes: 1
# push checkpoints to hub
hub_model_id: AlexHung29629/llama33_fix3
hub_strategy: end
# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
# Required to be true when used in combination with `push_dataset_to_hub`
hf_use_auth_token: true
# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.
val_set_size: 0

# The maximum length of an input to train with, this should typically be less than 2048
# as most models have a token/context limit of 2048
sequence_len: 2048
# Pad inputs so each step uses constant sized buffers
# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently
pad_to_sequence_len: true
# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'
sample_packing: false

# wandb configuration if you're using it
# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.
wandb_mode: disabled

# Tensorboard
use_tensorboard: true

# Where to save the full-finetuned model to
output_dir: ./step2-model


# Training hyperparameters

# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.
gradient_accumulation_steps: 1
# The number of samples to include in each batch. This is the number of samples sent to each GPU.
# Batch size per gpu = micro_batch_size * gradient_accumulation_steps
micro_batch_size: 1
num_epochs: 4
warmup_ratio: 0.0  # cannot use with warmup_steps
learning_rate: 2e-6
logging_steps: 1
save_strategy: epoch
#save_steps: 1000
saves_per_epoch: 1

# Whether to use gradient checkpointing. Available options are: true, false, "offload".
# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
gradient_checkpointing: true
# additional kwargs to pass to the trainer for gradient checkpointing
#gradient_checkpointing_kwargs:
#   use_reentrant: true

# Specify a scheduler and kwargs to use with the optimizer
lr_scheduler: cosine

optimizer: adamw_torch

# Specify weight decay
weight_decay: 0
# adamw hyperparams
adam_beta1: 0.9
adam_beta2: 0.95
adam_epsilon: 1e-8
# Gradient clipping max norm
max_grad_norm: 1

# Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
flash_attention: true
fsdp_final_state_dict_type: SHARDED_STATE_DICT
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD
  fsdp_backward_prefetch: BACKWARD_PRE
special_tokens:
  pad_token: <|finetune_right_pad_id|>
  eos_token: <|eot_id|>
# Deepspeed config path. e.g., deepspeed_configs/zero3.json
#deepspeed: /mnt/shared/twsc/alex/reasoning/zero3_bf16.json

# Seed
seed: 42
save_only_model: true

llama33_fix3

This model was trained from scratch on the train_v2.1.jsonl 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-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 8
• total_eval_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• num_epochs: 4.0

Training results

Framework versions

• Transformers 4.51.0
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1