SmolDataEnvs: fill empty file lists in 424 rows

#2
by thegovind - opened

files: [] appears in 424 rows: 398/5,000 train, 15/144 eval, and 11/250 test. Every row has data, but the GitHub single-program rollout builds its prompt's file list from files, so these prompts listed no file names.

Rows Before This PR
213 instruction has a placeholder; no file name appears anywhere in the row Put every staged file's flattened basename in files and the instruction's list
211 instruction names the files, sometimes with spaces or brackets, but files is empty Put those named, staged files in files; they can be a subset of all staged files

Only these rows change, in files and the 213 placeholder instructions. Schema, metadata, and parquet layout stay the same. No empty files list or placeholder remains.

In a check with base Qwen3-32B that listed every staged file, 38 affected tasks (15 eval, 23 from the GRPO training set) went from 0/304 to 79/304 correct. The companion GitHub PR has the details: https://github.com/adithya-s-k/FineEnvs/pull/21.

Testing

The file names from the 211 instruction-list rows were checked against staged files. On test, 11 rows change for the single-program prompt; instruction changes only for the 5 placeholder rows. Single-program test scores can move.

Publish this branch
This branch is in draft mode, publish it to be able to merge.

Sign up or log in to comment