base_model: Crystalcareai/Qwen-1.5-8x7B #this is the raw (random gated) model straight out of mergekit. Change this to "Crystalcareai/Qwen1.5-8x7b" for training SFT'd model.
model_type: Qwen2ForCausalLM #don't use HF auto config
tokenizer_type: Qwen2Tokenizer #don't use HF auto config
trust_remote_code: true


load_in_8bit: false
load_in_4bit: true #Mixtral models still chug vram in axolotl, so qlora is required at the moment.
strict: false


datasets:
  - path: Crystalcareai/MoD
        type: sharegpt
dataset_prepared_path: last_run_prepared #preprocess your dataset for easier vram: "python -m axolotl.cli.preprocess examples/Qwen/YOURCONFIG.yml"
val_set_size: 0.0
output_dir: ./qlora-out


model_config:
  output_router_logits: true


adapter: qlora
lora_model_dir:
sequence_len: 32768
sample_packing: true
pad_to_sequence_len: true


lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:


gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002  # anything from 2-5 is acceptable


train_on_inputs: false
group_by_length: false
bf16: true
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: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 4
debug:
deepspeed: deepspeed_configs/zero2.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens: