metadata

license: apache-2.0
library_name: peft
tags:
  - generated_from_trainer
base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
model-index:
  - name: qlora-out
    results: []
datasets:
  - totally-not-an-llm/ZorgonChat

See axolotl config

axolotl version: 0.4.0

base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: totally-not-an-llm/ZorgonChat
    type: alpaca
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./qlora-out

adapter: qlora
lora_model_dir:

sequence_len: 4096
sample_packing: false
pad_to_sequence_len: true

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

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 4
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0002

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: 10
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

qlora-out

This model is a fine-tuned version of TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T on the ZorgonChat dataset. It achieves the following results on the evaluation set:

Loss: 2.3466

Model description

Trained on a dataset of "alien language" chats to see if it will learn to talk in english. Prompt format is:

You are a helpful assistant., respond in Language: English

### Instruction:
{prompt}

### Response:

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
train_batch_size: 2
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 8
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 10
num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
3.9295	0.03	1	3.9073
3.5364	0.25	8	3.6199
3.263	0.5	16	3.1821
2.798	0.75	24	2.8962
2.7787	1.0	32	2.6773
2.5959	1.25	40	2.5506
2.4793	1.5	48	2.4955
2.5221	1.75	56	2.4613
2.4384	2.0	64	2.4055
2.295	2.25	72	2.3923
2.3943	2.5	80	2.3862
2.2398	2.75	88	2.3605
2.2693	3.0	96	2.3526
2.425	3.25	104	2.3471
2.2857	3.5	112	2.3468
2.2448	3.75	120	2.3451
2.1836	4.0	128	2.3466

Framework versions

PEFT 0.10.0
Transformers 4.40.0.dev0
Pytorch 2.1.2+cu118
Datasets 2.15.0
Tokenizers 0.15.0