Repository navigation
Learning from hindsight — research code and evidence
This is the first curated public release of Angel Raychev's September 2026
hindsight-learning implementation and research evidence.
The complete source archive includes the numerical learning machinery,
experiment coordinators, synthetic task data, compact historical results,
offline reanalysis, CPU correctness tests, and canonical report/source. The wheel
contains the Python library; use the source archive or repository for the full
research workflows and evidence.
Start with QUICKSTART.md:
- Regrade and rebuild the saved evidence using only Python's standard library.
- Exercise actual tiny/random-model optimization without downloading weights.
- Prepare new bounded real-model configurations only when you explicitly choose
to supply the model, GPU, storage and operator controls.
The original scientific core and historical evidence are unchanged. Release-only
coordinator repairs enforce configured qualification bounds and exact native
source/text admission. The expanded suite and CPU example passed a clean remote
package installation, and the downloadable source archive passed independent
extracted-tree checks. Provenance receipts pin those claims to their artifacts
and tested code commit.
This release does not rerun the scientific study or establish reliable
end-to-end learned-teacher improvement. Large probability tensors, fitted model
weights and the original private archive/history are not published. Application
timeouts are cooperative and do not terminate GPU billing.
Code is Apache-2.0 with retained upstream notices; see LICENSE and NOTICE.
The report is an authored technical manuscript, not an arXiv/conference submission.
Article
· Clean CPU verification
SHA256SUMS.txt covers every attached artifact.