4b.yml

# With base > Depth base + ChatML tokens replaced. 
base_model: IntervitensInc/Llama-3.1-Minitron-4B-Width-Base-chatml
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

## Including a few datasets i have not used before.
datasets:
  - path: Mielikki/Erebus-87k
    type: completion
    field: body
  - path: NewEden/Orion-Asstr-Stories-16K
    type: completion
    field: content
  - path: NewEden/Orion-LIT
    type: completion
    field: text
  - path: NewEden/Fujin-Cleaned-Final
    type: completion
    field: text
shuffle_merged_datasets: true
val_set_size: 0.0
output_dir: ./outputs

adapter:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:

sequence_len: 32768
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: true

wandb_project: tavbussy
wandb_entity:
wandb_watch:
wandb_name: Comp-attempt-1
wandb_log_model:


# Gradient clipping to keep grad norm down. Using the same as QwQ's max grad norm as distributions are similar(or not idr)
max_grad_norm: 0.2


hub_model_id: NewEden/Smol-pretrain
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true


gradient_accumulation_steps: 1
micro_batch_size: 8
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00001
weight_decay: 0.02

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_ratio: 0.1
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1

debug:
deepspeed: deepspeed_configs/zero3_bf16.json
fsdp:
fsdp_config:

special_tokens:
  pad_token: <|finetune_right_pad_id|>