Pythia Experiment 3: Gender bias

[haileyschoelkopf]

• Collect cooccurrence statistics on deduped and non-deduped Pile
• Cast WinoBias as a multiple choice classification task
• Convert corpus stats into Pandas dataset + visualize distribution
• Run WinoBias on all models
• Collect correlations between WinoBias scores and data statistics
• Create fine-tuning data points for insertion into last 3 ckpt retraining

[haileyschoelkopf]

The cooccurrence statistics were too uniformly noisy to derive meaningful conclusions from at large timescales.

Updated todos:

• train 1.3B on (M prn -> F prn) data but same data + order otherwise, for last 5k steps of training
• Run Winobias, AXG, crows_pairs, jigsaw on model before + after intervention
• ???

[haileyschoelkopf]

We aren't doing anything with the above described data, and ended up doing something else to fulfill this experiment.