initial README — paper scaffold
Browse files
README.md
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- temp-bench
|
| 5 |
+
- temporal-crosscoder
|
| 6 |
+
- sparse-autoencoder
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# temp-bench-models
|
| 10 |
+
|
| 11 |
+
Companion checkpoints for the **Temporal Crosscoders** paper (the `final` branch of
|
| 12 |
+
`chainik1125/temp_xc`). All checkpoints here were trained on the locked
|
| 13 |
+
two-architecture set:
|
| 14 |
+
|
| 15 |
+
- **TXC-base** = `txc_bare_antidead_t5` — vanilla TopK temporal crosscoder + tsae_paper anti-dead stack.
|
| 16 |
+
- **TXC-pro** = `phase5b_subseq_h8` — subseq encoder + matryoshka H8 + multi-distance contrastive.
|
| 17 |
+
|
| 18 |
+
Plus baselines used in the paper: TopK-SAE, T-SAE (paper config), TFA, MLC, SAE-arditi (C6 only).
|
| 19 |
+
|
| 20 |
+
## Layout
|
| 21 |
+
|
| 22 |
+
```
|
| 23 |
+
<run_id>/
|
| 24 |
+
config.json arch + training config
|
| 25 |
+
model.safetensors weights
|
| 26 |
+
manifest.json run-level metadata (component, seed, k, metrics)
|
| 27 |
+
```
|
| 28 |
+
|
| 29 |
+
`<run_id>` follows `<component>_<arch>_<seed>_<short-hash>` from the
|
| 30 |
+
paper's run-id contract. Browse `purified/checkpoints/manifest.jsonl` in
|
| 31 |
+
the repo for the canonical index.
|
| 32 |
+
|
| 33 |
+
## Privacy
|
| 34 |
+
|
| 35 |
+
Private during paper development. Will flip to public when the paper draft stabilises.
|