RichardErkhov/Dans-DiscountModels_-_Dans-Instruct-Mix-8b-ChatML-V0.0.1-gguf

base_model: Dans-DiscountModels/Meta-Llama-3.1-8B-ChatML
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

trust_remote_code:

# wandb configuration
wandb_project: l3.1-8b-dans-instruct
wandb_watch:
wandb_run_id:
wandb_log_model: 

# where to save the finished model to
output_dir: ./l3.1-8b-dans-instruct

# dataset settings (local or huggingface repo)
datasets:
  - path: PocketDoc/Dans-MemoryCore-CoreCurriculum-Small
    type: sharegpt
    conversation: chatml
  - path: AquaV/Energetic-Materials-Sharegpt
    type: sharegpt
    conversation: chatml
  - path: AquaV/Chemical-Biological-Safety-Applications-Sharegpt
    type: sharegpt
    conversation: chatml
  - path: PocketDoc/Dans-Mathmaxx
    type: sharegpt
    conversation: chatml
  - path: PocketDoc/Dans-Benchmaxx
    type: sharegpt
    conversation: chatml
  - path: PocketDoc/Dans-Codemaxx
    type: sharegpt
    conversation: chatml
  - path: PocketDoc/Dans-Taskmaxx
    type: sharegpt
    conversation: chatml
  - path: PocketDoc/Dans-ASCIIMaxx-Wordart
    type: sharegpt
    conversation: chatml
  - path: PocketDoc/Dans-Prosemaxx
    type: sharegpt
    conversation: chatml
  - path: PocketDoc/Dans-Toolmaxx
    type: sharegpt
    conversation: chatml

chat_template: chatml

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

dataset_prepared_path: ./l3.1-8b-dans-instruct-data
val_set_size: 0.03

lora_model_dir: 

sequence_len: 8192

# use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'
sample_packing: true
eval_sample_packing: true

# you can set these packing optimizations AFTER starting a training at least once.
# The trainer will provide recommended values for these values.

pad_to_sequence_len: true

#rope_scaling:
  #type:  # linear | dynamic
  #factor:  # float (2 for 2x)

adapter: # blank for full finetune
lora_r: 64
lora_alpha: 64
lora_dropout: 0.2
lora_target_linear: True
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
  - gate_proj
  - down_proj
  - up_proj
lora_modules_to_save:
  - embed_tokens
  - lm_head
lora_fan_in_fan_out:

gradient_accumulation_steps: 32
micro_batch_size: 1
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.0000015
cosine_min_lr_ratio: 

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

gradient_checkpointing: unsloth
early_stopping_patience:
resume_from_checkpoint:
auto_resume_from_checkpoints: false
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 15
eval_steps: 25
# save_steps: 100
saves_per_epoch: 3
debug: false
deepspeed:
weight_decay: 0.1
fsdp:
fsdp_config:


special_tokens:
  pad_token: <|finetune_right_pad_id|>
  eos_token: <|im_end|>