stillerman/mtg-aurora

See axolotl config

axolotl version: 0.4.0

base_model: aurora-m/aurora-m-v0.1 # this can be swapped for mdel model when the model is released
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
is_llama_derived_model: false

load_in_8bit: false # when this is true inference quality is terrible
load_in_4bit: false
strict: false

datasets:
  - path: /workspace/axolotl-mdel/mtg.txt # change this to where your dataset is
    type: completion # change this to 'alpaca' if you are using alpaca formatting

lora_modules_to_save:
  - embed_tokens
  - lm_head

dataset_prepared_path:
val_set_size: 0.05
output_dir: ./lora-out

sequence_len: 4096 # this can be tweaked for efficiency
sample_packing: true
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: mtg-aurora-experiement # give this a name
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 2 # this can be tweaked for efficiency
micro_batch_size: 1 # this can be tweaked for efficiency
num_epochs: 1 # this can be experimented with
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: false # when this is true, inference quality is terrible
s2_attention:

warmup_steps: 10 # this can be tweaked for efficiency
evals_per_epoch: 10 # this can be tweaked for efficiency
eval_table_size:
eval_table_max_new_tokens: 128
saves_per_epoch: 1 
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: "<|endoftext|>"
  eos_token: "<|endoftext|>"

lora-out

This model is a fine-tuned version of aurora-m/aurora-m-v0.1 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.7945

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
4.2833	0.0	1	4.0839
2.1947	0.1	25	1.9886
1.2659	0.21	50	1.1937
1.0662	0.31	75	1.0060
0.9538	0.41	100	0.9172
0.9232	0.52	125	0.8603
0.8546	0.62	150	0.8237
0.8223	0.73	175	0.8049
0.8546	0.83	200	0.7979
0.8995	0.93	225	0.7945

Framework versions

• PEFT 0.7.2.dev0
• Transformers 4.37.0
• Pytorch 2.1.2+cu118
• Datasets 2.16.1
• Tokenizers 0.15.0