Blame for README.md

---
pretty_name: Temporal Crosscoders paper artifacts
tags:
  - interpretability
  - sparse-autoencoders
  - temporal-crosscoders
  - reproducibility
---

## Temporal Crosscoders artifacts

Public numerical inputs and archived outputs for [Temporal Crosscoders](https://github.com/chainik1125/temporal-crosscoders).

From a checkout of the code repository, run `uv run python figures.py all`.
The code pins this dataset to a commit and verifies every downloaded file's SHA-256.
The figure inputs and six archived appendix images total approximately 5.6 MB;
plotting does not download model weights or require a GPU or API key.

- `figure-data/`: frozen 2026-11-07 camera-ready numerical inputs, raw source records,
  per-seed values and provenance. The current main plots exclude TXC-pro.
- `reference-figures/`: six original appendix images whose exact input records
  have not been recovered. The reproduction command identifies these as archived,
  not regenerated from data.
- `anonymous-release/artifacts.tar.gz`: all tracked non-code experiment artifacts
  from the anonymous-release Git snapshot, with their original relative paths.
- `anonymous-release/manifest.json`: per-file sizes and checksums for that archive.

The two large upstream repositories are also public:
[trained dictionary checkpoints](https://huggingface.co/han1823123123/temp-bench-models/tree/1e8d750f61b82cabf1981b098070cb334e5d7fb9)
and [activation caches and data](https://huggingface.co/datasets/han1823123123/temp-bench-data/tree/3f7c4a0b9bdf3480936d2db462cc832bab783c11).
They retain their upstream layout. Their presence does not establish that every historical run's
checkpoint or raw generation is available; see the figure-specific provenance and missing-data records.

The original snapshot is historical: its measurements must not be substituted for the
current camera-ready figures without checking the evaluation recipe. Base models and datasets
retain their upstream terms. No base language-model weights are redistributed here.