base_model: IntervitensInc/Llama-3.1-Minitron-4B-Width-Base-chatml
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_4bit: false
strict: false

datasets:
#  - path: anthracite-core/c2_logs_32k_mistral-v3_v1.2
#    type: sharegpt
#    conversation: chatml
  - path: ./datasets/c2_deduped_32k_mistral-v3_tok_deanon_dsclean_1.2.jsonl
    type: sharegpt
    conversation: chatml
#  - path: anthracite-org/kalo-opus-instruct-22k-no-refusal
#    type: sharegpt
#    conversation: chatml
  - path: ./datasets/opus-instruct-22k-no_refusals.jsonl
    type: sharegpt
    conversation: chatml
#  - path: lodrick-the-lafted/kalo-opus-instruct-3k-filtered
#    type: sharegpt
#    conversation: chatml
  - path: ./datasets/kalo-3k-filtered.jsonl
    type: sharegpt
    conversation: chatml
#  - path: anthracite-org/nopm_claude_writing_fixed
#    type: sharegpt
#    conversation: chatml
  - path: ./datasets/claudewritingNopm.jsonl
    type: sharegpt
    conversation: chatml
#  - path: anthracite-org/kalo_opus_misc_240827
#    type: sharegpt
#    conversation: chatml
  - path: ./datasets/kalo_opus_misc_240827.jsonl
    type: sharegpt
    conversation: chatml
#  - path: anthracite-org/kalo_misc_part2
#    type: sharegpt
#    conversation: chatml
  - path: ./datasets/kalo_misc_part2.jsonl
    type: sharegpt
    conversation: chatml
#  - path: NewEden/Claude-Instruct-5K
#    type: sharegpt
#    conversation: chatml
  - path: ./datasets/5k.jsonl
    type: sharegpt
    conversation: chatml

#chat_template: chatml
shuffle_merged_datasets: true
#default_system_message: "You are an assistant that responds to the user."
dataset_prepared_path: ./magnum-22b-data
val_set_size: 0.0
output_dir: ./22b-fft-out

sequence_len: 16000
sample_packing: true
pad_to_sequence_len: true


wandb_project: 4bmagnum
wandb_entity:
wandb_watch:
wandb_name: 4magnum
wandb_log_model:

gradient_accumulation_steps: 32
micro_batch_size: 1
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: 40
evals_per_epoch:
eval_table_size:
eval_max_new_tokens:
saves_per_epoch: 2
debug:
#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
deepspeed: /workspace/axolotl/deepspeed_configs/zero2.json
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|finetune_right_pad_id|>