lightonai/SASS-Post-Training-Dataset

SASS training data

Unfiltered DeepSeek SFT trajectories for four search/language setups, plus the shared multilingual RL task corpus. This repository contains no weighted or localization-filtered SFT subsets.

Configuration	Split	Rows
grep-python-unfiltered	sft	14,500
colgrep-python-unfiltered	sft	14,500
grep-multi-unfiltered	sft	14,592
colgrep-multi-unfiltered	sft	14,028
rl-multi	sft	10,899

from datasets import load_dataset
sft = load_dataset("lightonai/SASS-Data", "colgrep-python-unfiltered", split="sft", token=True)
rl = load_dataset("lightonai/SASS-Data", "rl-multi", split="rl", token=True)

The rl-multi configuration contains the same 10,899 source tasks used for both GREP and ColGREP RL: problem statements, repositories, base commits and gold file/line spans, not policy rollouts. It replaces the duplicated rl splits formerly nested under grep and colgrep. Those two ambiguous configurations previously exposed weighted multilingual SFT and have been removed. RL files are unchanged; SFT-only columns in their shared nullable schema are null.

Provenance

SFT source revisions and checksums are recorded in release-manifest.json. Unfiltered source rows are retained without applying localization-quality filters or turn weighting. The original source repositories still provide the weighted and filtered variants used for the paper's existing checkpoints; removing those copies here does not alter any model or source repository.

RL comes from lightonai/code-search-corpus-RL-multi at revision 47af6636335e5515be3e9b50cd673ca1940c2d79 (33 repositories, eight languages). Both historical RL environment logs confirm that indexed_only=true retained all 10,899/10,899 source tasks. All source tasks were eligible; eligibility does not imply that every task was sampled during the finite training run.

Python generation preserves its archived harness lifecycle (ten search turns, a reminder, two finish-only turns and rescue). Multilingual v2 uses a hard ten-response cap and stricter protocol cleaning. Matching unfiltered counts does not remove these existing protocol differences. ColGREP Python unfiltered trajectories retain their strict teacher prompt. protocols/ contains historical student protocols; the packaged ColGREP protocol differs from the printed Appendix A.6 prompt. Source/model terms continue to apply.

Python count-matching extension

Both Python setups now contain 14,500 unfiltered trajectories. Added trajectories were generated with DeepSeek-V4-Flash-0731 using the frozen Python harness, temperature 0.3, thinking disabled, and existing decontaminated tasks screened against original Python tasks and evaluation overlap. Candidate tasks, generation manifests, frozen runtime and added-shard hashes are retained under python-balanced-20261007/. These additions postdate the reported paper experiments and were not used by the existing checkpoints.