sashaboguraev/pythia-1b-ppt-control_music_steps100_1b-seed1024

Stage-1 final for the no-warmup retrain

Retrain without warmup, matching the other 187 1B runs (linear decay from 3e-4, no warmup) on transformers 5.4.0. Replaces the Aug 2026 run whose 1000-step warmup made it ~0.27 nats better than every other cell. Prior revisions retain the superseded weights.

Retrain without warmup, matching the other 187 1B runs (linear decay from 3e-4, no warmup) on transformers 5.4.0. Replaces the Aug 2026 run whose 1000-step warmup made it ~0.27 nats better than every other cell. Prior revisions retain the superseded weights.

Upload folder using huggingface_hub

Upload tokenizer

Upload GPTNeoXForCausalLM

initial commit