Raw data
Every number behind the figures, to check or explore on your own. The table of versions has one row per model file. The test answers view lists each scored answer separately. The per-digit view shows how sure each version was of each of the first 1,000 digits of π.
One row per model version, 74 in all: the 64 versions of the main study plus 10 rebuilds of unsloth's recipe from follow-up 3. Click a column header to sort, and hover over it to see what it measures. Shading marks where each value sits among the rows shown, darker for larger. It shows magnitude, not better or worse.
What every column means
Every scored answer, one row each: three wordings × three constants for E1, 40 strings for E3, 30 for E4b, 8 planted digits for E5 and π, e and √2 for E6. Error positions use S for a swapped digit, I for an inserted one and D for a dropped one, followed by the position in the expected text. The model's own words aren't published here, only their scores.
Teacher forcing on the E2 π text. Each row is one version, and each column one digit position: position n is the digit that follows n correct ones, so a free recitation that stops at 813 went wrong at position 813. Hover over a cell for its numbers, and click one to list the positions around it. How teacher forcing works.