Lora finetuning on Wikipedia-10, applying counter factual data augmentation (CDA)

Dataset: Wikipedia-10
Target modules = ["q_proj", "k_proj", "v_proj", "dense", "fc1", "fc2"]

{
    "epoch": 0.22326645805206993,
    "total_flos": 1.04520375926784e+18,
    "train_loss": 2.3526821104288103,
    "train_runtime": 66080.1443,
    "train_samples": 286653,
    "train_samples_per_second": 0.969,
    "train_steps_per_second": 0.03
}

Training script: https://github.com/ao9000/bias-bench/blob/main/experiments/run_clm.py

Downloads last month: -

Safetensors

Model size

3B params

Tensor type

BF16