Live status / Language models
Language models
Two open language models, each tuned for one job: a referee that checks a paper's claims by recomputing them, and a cleaner that removes the habits of machine-written prose without changing a number.
-
The referee: teacher traces under way
RunningThe referee learns from worked examples. The base model answers each training question with its tools, code checks every answer, and only answers that pass are kept for tuning. The first long run closed 2,026 of the 6,266 question episodes before its time ran out and kept 175 checked traces. Finished episodes are kept, the step budget each episode gets has been fixed, and the run picks up where it stopped.
Figure 1. Episodes closed, by kind, in the first long run of the teacher replay, which ended at its time limit on 12 October. A planted fault is an error put into a paper on purpose for the referee to find. The synthetic conversations all closed. Which traces were kept
Of the 175 traces kept, 28 are about claims, 25 about planted faults, 63 about papers with no fault and 59 are synthetic. The most common reason a trace was dropped was a verdict that cited a number the tools never returned, 306 times.
-
The cleaner: timing run done, fine-tune next
NextThe cleaner rewrites machine-sounding prose and must keep every number. A short timing run on the first base model finished on 12 October: 55 percent of passages came out clean, with a preservation score of 0.80. The full fine-tune comes next; an earlier form already serves at home.
-
Downloads and housekeeping
Base models and data come down to the supercomputer on transfer nodes, and finished runs are archived before their folders are cleared.