Random 160M β€” uniform WSD v2

Full research checkpoints (model, AdamW, scheduler, rank RNG and injection counters). These are torch training states, not Transformers from_pretrained exports.

Priority: MCG, M, C, G at r32, then r16/r8/r4/r2/r1. Each condition targets 16B stable tokens. Each 1B anchor has an independent 100M background-only cooldown. Historical constant-LR controls are retained; new controls are disabled. Funding may cover only a prefix of this queue. No missing condition is implied complete.

512-token non-generation evaluation uses frozen 1536 targets (512 each M/C/G), plain and triggered. Raw NLL and trigger effects are published after each complete trajectory. Shared recipe-matched Clean stable/cooldown baselines are scored on the identical bank. They use a different numerical training execution and are not a bitwise reproduction. Historical controls have no matched Clean adjustment.

Results: matched Clean-adjusted JSON, 512-token primary CSV, and Clean primary CSV. Missing conditions remain missing. See the training, evaluation and handoff guide for completed trajectories, frozen identities and execution provenance.

Source: https://github.com/294-288-pretrain-interaction/pretrain_interaction/tree/6d3bdfc380c6e56cff3bd010e0a844a650991865 Branch: yihan/random-wsd-v2. Checkpoint paths: uniform-wsd-v2/160m//.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support