jsaurabh/weight-space-lab-artifacts

Weight Space Lab: model artifacts

Checkpoints, weight edits and activation arrays for Weight Space Lab. Code and concise experiment reports live on GitHub; raw generation logs and API transcripts are archived here. artifact_inventory.json records exact SHA-256 checksums and provenance.

Contents

• results/reference/{parity,addition}/weights.pt: guided transformer constructions. Both exhaustively solve their documented finite domains.

• results/discovery/astra-initial/{parity,addition}/weights.pt: actual independent Astra-generated numerical constructions. Both exhaustively solve their documented finite domains. Files contain a PyTorch state dictionary, architecture configuration, task and provenance; load with torch.load(..., weights_only=True).

• third_party/qwen_p50c/edit.npz: the authors' unmodified published sparse Qwen edit, with source manifest, provenance and CC BY 4.0 attribution.

• results/qwen/blind-discovery/selected.safetensors: the independently selected Qwen candidate. This contains only two edited tensors, not a full model. The blind search did not improve performance: base 5/64 versus candidate 3/64; both finished correctly 3/64 times and hit the token cap 61/64 times.

• results/qwen/blind-discovery/artifacts/*.npz: recorded activation arrays from that independent search, without object pickles.

• results/comparison/qwen-astra-vs-gpt55-20261103-interrupted/gpt55/artifacts/*.npz: two activation captures from an interrupted infrastructure trial, excluded from the editor comparison. See the interruption report.

• results/comparison/qwen-astra-vs-gpt55-20261103/{gpt55,astra}/selected.safetensors: independent frozen patches from the matched ten-feedback-call Qwen pilot. GPT-5.5 changes one matrix; Astra changes 24. On 32 unseen MATH-500 questions with two samples each, base scored 7/64, Astra 11/64, and GPT-5.5 13/64. This is one attempt per editor; Astra required declared infrastructure recoveries. See the full report and limits. Associated artifacts/*.npz files contain two activation captures per editor.

• results/comparison/qwen-astra-vs-gpt55-20261103/recovery_001/launcher_repair/current_checkpoint.safetensors: an intermediate Astra checkpoint retained as recovery evidence. This is not either final selected repair.

Reproduce

Clone the GitHub repository and run python3 experiments/fetch_artifacts.py. This restores the model artifacts to their expected paths and verifies each checksum against a GitHub manifest pinned to a Hugging Face commit. The GitHub README links all completed experiments and replay instructions. Fetching artifacts makes no editor API calls and starts no GPU instance.

The Qwen patches require Qwen/Qwen3.5-0.8B revision 2fc06364715b967f1860aea9cf38778875588b17. Full base weights are available from the original model repository and are not duplicated here. The independent patch can be materialized with experiments/blind/replay.py from GitHub.

Attribution

The published numerical diff is by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, Sparks of Weight-Space Programming (2026), and retains its included CC BY 4.0 attribution. The modified Qwen tensors derive from the Qwen team's Apache 2.0 model; its license and the modification notice are included alongside them. The tiny checkpoints were constructed for this project without gradient descent. See each GitHub report for the distinction between guided construction, independent discovery and replay.

Archived experiment records

The records/ archives hold raw API exchanges, generated answers, diagnostics, source snapshots and recovery records. GitHub keeps code and concise reports. python3 experiments/fetch_artifacts.py restores these records under the ignored results/ directory, together with the separately stored checkpoints and activation arrays. Archive hashes and sizes are pinned in GitHub’s model_artifacts.json. Group-specific downloads also restore the relevant records.