llama2-7b-auto-improving-boolq

<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.4.0

# This file is used by the training script in train.ipynb. You can read more about
# the format and see more examples at https://github.com/OpenAccess-AI-Collective/axolotl.
# One of the parameters you might want to play around with is `num_epochs`: if you have a
# smaller dataset size, making that large can have good results.

base_model: meta-llama/Llama-2-7b-hf
base_model_config: meta-llama/Llama-2-7b-hf
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

load_in_8bit: true
load_in_4bit: false
strict: false

datasets:
  - path: ./resources/train_aug.jsonl
    type: alpaca
dataset_prepared_path: ./resources/last_run_prepared
val_set_size: 0.05
output_dir: ./models/auto-improving-run

sequence_len: 4096
sample_packing: 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:

# This will report stats from your training run to https://wandb.ai/. If you don't want to create a wandb account you can comment this section out.
wandb_project: google-boolq
wandb_entity:
wandb_watch:
wandb_run_id: auto-improving-run
wandb_log_model:


gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 5
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
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

warmup_steps: 10
eval_steps: 20
save_steps: 60
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

</details><br>

models/auto-improving-run

This model is a fine-tuned version of meta-llama/Llama-2-7b-hf on the google/boolq dataset with a research platform that iterates on the model inaccuracies, gets refined by expert, and re-performs training. It achieves the following results on the evaluation set:

• Loss: 0.3435

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: 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: 5

Training results

Framework versions

• PEFT 0.9.0
• Transformers 4.40.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.18.0
• Tokenizers 0.15.0

Evaluation

These are the metrics reported on the test data (10% of boolq)

model='auto-improving-llama' accuracy=0.8629969418960245 avg_time=0.044935779816513415 avg_cost=1.6101987767584347e-05