axolotl-ai-co/numina-70b-ep3-lr1e-5-sft-lora

See axolotl config

axolotl version: 0.6.0

base_model: meta-llama/Llama-3.1-70B
tokenizer_type: AutoTokenizer

strict: false

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.spectrum.SpectrumPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
  # liger_cross_entropy: true
  # liger_fused_linear_cross_entropy: true

  # torch_compile: true

dataloader_prefetch_factor: 8192
dataloader_num_workers: 8
dataloader_pin_memory: true

dataset_processes: 16

chat_template: llama3
datasets:
  - path: AI-MO/NuminaMath-CoT
    type: chat_template
    split: train
    field_messages: messages
    message_field_content: content
    message_field_role: role


dataset_prepared_path: /workspace/data/axolotl-artifacts/last_run_prepared
val_set_size: 0.0
output_dir: /workspace/data/axolotl-artifacts/outputs/llama3_1-70b-finetome
save_safetensors: false  # saving final sharded dict may not work with safetensors

wandb_project: numina-kd-experiment
wandb_entity: axolotl-ai

adapter: lora
  # peft_use_dora: true
lora_r: 512
lora_alpha: 1024
lora_dropout: 0.05
lora_target_linear: true
lora_modules_to_save:
  - embed_tokens
  - lm_head

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

gradient_accumulation_steps: 1
# 8x Node can support a batch size of up to 3
micro_batch_size: 8
num_epochs: 3
optimizer: optimi_adamw
lr_scheduler: cosine
learning_rate: 1.0e-5

train_on_inputs: false
group_by_length: false
bf16: true
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: true
logging_steps: 1
flash_attention: true

warmup_steps: 40
saves_per_epoch: 1
weight_decay: 0.1
deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_params.json
special_tokens:
  pad_token: <|finetune_right_pad_id|>
  eos_token: <|eot_id|>

workspace/data/axolotl-artifacts/outputs/llama3_1-70b-finetome

This model is a fine-tuned version of meta-llama/Llama-3.1-70B on the AI-MO/NuminaMath-CoT 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: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 64
• total_eval_batch_size: 64
• optimizer: Use OptimizerNames.ADAMW_HF 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: 40
• num_epochs: 3

Training results

Framework versions

• PEFT 0.14.0
• Transformers 4.48.1
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0