ilyadorosh/love

See axolotl config

axolotl version: 0.5.0

base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
model_type: AutoModelForCausalLM
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

plugins:
  - axolotl.integrations.liger.LigerPlugin

liger_glu_activation: true
liger_rms_norm: true
liger_layer_norm: true

load_in_4bit: true
adapter: qlora
lora_r: 32
lora_alpha: 64
lora_target_modules:
  - self_attn.q_proj
  - self_attn.k_proj
  - self_attn.v_proj
  - self_attn.o_proj
  - shared_expert.gate_proj
  - shared_expert.up_proj
  - shared_expert.down_proj
  # - experts.gate_projs.[0-9]+$  # optionally train the moe experts
  # - experts.up_projs.[0-9]+$
  # - experts.down_projs.[0-9]+$
lora_modules_to_save:
   - lm_head  # needed if modifying vocabulary
   - embed_tokens

lora_mlp_kernel: true
lora_qkv_kernel: true
lora_o_kernel: true

chat_template: chatml
datasets:
  - path: mlabonne/FineTome-100k
    type: chat_template
    split: train[:20%]
    field_messages: conversations
    message_field_role: from
    message_field_content: value

dataset_prepared_path: last_run_prepared
val_set_size: 0.0
output_dir: ./outputs/out

sequence_len: 4096  # up to 8k will work on a single H100
sample_packing: true


gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_8bit
lr_scheduler: cosine
learning_rate: 1e-4

bf16: true
tf32: true

torch_compile: true
flex_attention: true
flex_attn_compile_kwargs:
  dynamic: false
  mode: max-autotune-no-cudagraphs

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false

logging_steps: 1
warmup_ratio: 0.1
evals_per_epoch: 1
saves_per_epoch: 1

weight_decay: 0.0
special_tokens:
  pad_token: <|finetune_right_pad|>
  eos_token: <|eot|>

# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

outputs/out

This model is a fine-tuned version of TinyLlama/TinyLlama-1.1B-Chat-v1.0 on the None 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: 0.0001
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_8BIT 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: 353
• num_epochs: 1

Training results

Framework versions

• PEFT 0.13.2
• Transformers 4.46.1
• Pytorch 2.3.1+cu121
• Datasets 3.0.1
• Tokenizers 0.20.3