PygTesting/pyg3old_qlora_adapter

See axolotl config

axolotl version: 0.4.1

base_model: mistralai/Mistral-Nemo-Base-2407
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

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

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

loraplus_lr_ratio: 2  

chat_template: chatml

datasets:
  - path: /home/austin/disk2/axolotl_data/fixed_pyg3.jsonl
    type: sharegpt
    conversation: chatml

dataset_prepared_path: /home/austin/disk2/axolotl_data/data_tokenized
val_set_size: 0.01
output_dir: /home/austin/disk2/axolotl_storage/pyg3_qlora_2e-4

sequence_len: 8192
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

wandb_project: pyg3-qlora
wandb_entity:
wandb_watch:
wandb_name: 1e-5
wandb_log_model:
 
 #unsloth_cross_entropy_loss: true

gradient_accumulation_steps: 1
micro_batch_size: 3
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 100
evals_per_epoch: 10
eval_table_size:
saves_per_epoch: 10
debug:
deepspeed: deepspeed_configs/zero2.json
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:
  pad_token: </s>

home/austin/disk2/axolotl_storage/pyg3_qlora_2e-4

This model is a fine-tuned version of mistralai/Mistral-Nemo-Base-2407 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.8024

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: 3
• eval_batch_size: 3
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 24
• total_eval_batch_size: 24
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.8656	0.0006	1	1.1181
1.5716	0.1004	175	0.8479
1.6573	0.2008	350	0.8308
1.8387	0.3012	525	0.8230
1.5855	0.4016	700	0.8167
1.7139	0.5020	875	0.8123
1.5684	0.6024	1050	0.8087
1.6986	0.7028	1225	0.8055
1.6505	0.8032	1400	0.8035
1.6028	0.9036	1575	0.8024

Framework versions

• PEFT 0.12.0
• Transformers 4.44.0
• Pytorch 2.4.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1