ReLaDrift trained models

Checkpoints of ReLaDrift: Recurrent Language Generation via Drifting for Test-Time Scaling.
Code: Yuki-Ots/ReLaDrift.
Training data: CollinsStreet/ReLaDrift-data.

Path	Contents
gsm8k/reladrift/	15 conditions with raw and EMA weights: base (ReLaDrift), r1ns (Drifting Model) and ablations; 250k updates (tau1, v1pq, v1pq_n4: 40k)
gsm8k/baselines/	AR, S-FLM, Duo, MDLM and FMLM+, 250k updates (upstream Lightning checkpoints)
gsm8k/decoder/	Latent-to-token decoders and the random embedding table of the rndtab ablation
sudoku/reladrift/	c{01..11}_seed{0,1,2}.pt: 11 conditions (c01 ReLaDrift, c04 Drifting Model and ablations), 20k updates
sudoku/baselines/	{ar,sfm,duo,mdlm,fmlm_plus}_seed{0,1,2}.pt: AR, S-FLM, Duo, MDLM and FMLM+, 20k updates
sudoku/frm/	frm_seed{0,1,2}.pt: FRM, 20k updates
sudoku/dfm/	{base,ng4,ng8,ng16}_seed{0,1,2}.pt: DFM with four settings, 20k updates

Sudoku files hold the full training state. SHA256SUMS lists every file under gsm8k/ and sudoku/. In the code repository, bash scripts/data/download_checkpoints.sh downloads and verifies them, and the scripts in scripts/sample/ evaluate them.

License

The checkpoints are released under the Apache License 2.0. The AR, S-FLM, Duo, MDLM and FMLM+ baselines were trained with code imported from S-FLM and posterior-refinement (FMLM+), both distributed under the Apache License 2.0.