|
Download PUBLICATION_INDEX.md from gifoe/loop: direct link, hf CLI and curl.
- Browser
- Download file 1.89 kB
-
https://huggingface.co/datasets/gifoe/loop/resolve/main/PUBLICATION_INDEX.md
- Command line
-
hf download hf://datasets/gifoe/loop/PUBLICATION_INDEX.md
-
curl -L -o PUBLICATION_INDEX.md https://huggingface.co/datasets/gifoe/loop/resolve/main/PUBLICATION_INDEX.md
1.89 kB
| # Ouro loop experiment publication | |
| This public dataset is being populated. Its existence does **not** mean that all experiments or raw activations have been uploaded. Files are added incrementally and verified against source SHA-256 and size; do not infer missing files are complete from a manifest or a successful test upload. | |
| The reviewed, in-progress Chinese report is [FINAL_RESEARCH_REPORT_reviewed_20261112.md](FINAL_RESEARCH_REPORT_reviewed_20261112.md). It distinguishes natural-trajectory observations from activation interventions, records negative and incomplete outcomes, and describes the historical endpoint discrepancy. The QKV train analysis has since progressed; its full report and later validation/test analysis will be added after scientific review. | |
| The current code and report snapshot is on the GitHub [research branch](https://github.com/Gifoe/looped-transformer-stage1-reproduction/tree/research/ouro-latent-mechanism-20261107). Its curated export identifies source commit `1d1f8c0fcaa7aadf5aece910b159f2f920ebb8f0`; it does not claim to reproduce all of that commit's unreviewed history. A first code and research-plan archive is available at [publication/ouro-publication-small-20261111.tar.gz](publication/ouro-publication-small-20261111.tar.gz), SHA-256 `14e31257cf45cf1a3e09f5ef069278592e27f0e898f5a3f4dd00e9a958a5128b`; it is a partial snapshot, not the complete results. | |
| Experiment files are being published under `results/` with project-relative paths. Raw tensor, score and metadata files are tracked against a fixed per-file manifest; upload tests under `transport-test/` can duplicate those results and must not be counted again as unique scientific data. A final complete, verified per-file inventory and restoration instructions will be supplied once transfers finish. Model checkpoints and authentication material are not part of this dataset. | |