base_model: meta-llama/Llama-3.2-3B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

load_in_8bit: false
load_in_3bit: false
strict: false

datasets:
 - path: NewEden/Gryphe-Sonnet3.5-Charcard-Roleplay-unfiltered
   type: sharegpt
   conversation: llama3
 - path: Gryphe/ChatGPT-4o-Writing-Prompts
   type: sharegpt
   conversation: llama3
 - path: Epiculous/SynthRP-Gens-v1.1-Filtered-n-Cleaned
   type: sharegpt
   conversation: llama3
 - path: Epiculous/Synthstruct-Gens-v1.1-Filtered-n-Cleaned
   type: sharegpt
   conversation: llama3
 - path: NewEden/Gryphe-Sonnet-3.5-35k-Subset
   type: sharegpt
   conversation: llama3
 - path: anthracite-org/kalo_opus_misc_240827
   type: sharegpt
   conversation: llama3
 - path: anthracite-org/kalo_misc_part2
   type: sharegpt
   conversation: llama3
 - path: NewEden/Claude-Instruct-5K
   type: sharegpt
   conversation: llama3
 - path: anthracite-org/kalo-opus-instruct-22k-no-refusal
   type: sharegpt
   conversation: llama3
chat_template: llama3
shuffle_merged_datasets: true
#default_system_message: "You are an assistant that responds to the user."
dataset_prepared_path: ./holland-3b-data
val_set_size: 0.0
output_dir: ./3b-fft-out

sequence_len: 32768
sample_packing: true
pad_to_sequence_len: true

wandb_project: Holland-Mag-3B-LLama
wandb_entity:
wandb_watch:
wandb_name: attempt-llama3-01
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 4
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.000005

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

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

warmup_steps: 10
evals_per_epoch:
eval_table_size:
eval_max_new_tokens:
saves_per_epoch: 2
debug:
deepspeed:
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|finetune_right_pad_id|>