Remove the GLM-5.3 rollout and teacher comparison while retaining both task variants and unchanged GLM-5.2 data. Restore the scaffold-only dataset configs and document evaluation safeguards: anti-hack prompts, Git information removal, and restricted agent internet access.
- trajectories/legoflow-selfmade-*.jsonl renamed to legoflow-swe-*.jsonl (server-side LFS copy, old paths removed; .gitattributes and config data_files updated)
- README rewritten against the latest blog: three-benchmark results, reasoning-depth and GLM-5.3 teacher sections, curator-selfmade relabelled LegoFlow-SWE
- new figure set added under figures/
Name GLM-5.2 trajectory subsets explicitly as allow-hack
Rename the OpenHands SDK and OpenCode configs to include GLM-5.2 and allow-hack, and update loading instructions. Preserve the existing data-file mappings and OpenCode features.
Add GLM-5.3 anti-hack trajectories and visible reward columns
Publish 4,730 GLM-5.3/OpenHands SDK anti-hack conversations (1,346 reward=1; 3,384 reward=0, including 31 missing rewards mapped to zero). Add top-level reward to both existing GLM-5.2 JSONLs without altering their original fields. Document task-version provenance and retain the OpenCode tool parameter schemas with explicit JSON features.
Complete anti-hack task release with original binary fixtures
