Initial README — Phase 7 ckpts on Gemma-2-2b base
Browse files
README.md
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- sparse-autoencoder
|
| 5 |
+
- temporal-crosscoder
|
| 6 |
+
- gemma-2-2b
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# txcdr-base — Temporal Crosscoder checkpoints (Phase 7, Gemma-2-2b base)
|
| 10 |
+
|
| 11 |
+
This repo contains SAE / TXC / MLC / TFA checkpoints **trained on
|
| 12 |
+
`google/gemma-2-2b` (base, NOT IT)** as part of Phase 7 of the
|
| 13 |
+
Temporal Crosscoders project.
|
| 14 |
+
|
| 15 |
+
All ckpts in this repo were trained at:
|
| 16 |
+
|
| 17 |
+
- Subject model: `google/gemma-2-2b` (base).
|
| 18 |
+
- Anchor layer: 0-indexed L12 (≈ 50% model depth) — matches T-SAE
|
| 19 |
+
(Ye et al. 2025) and TFA (Lubana et al. 2025).
|
| 20 |
+
- Sparsity: `k_win = 500` global TopK on `d_sae = 18432` (with one
|
| 21 |
+
paper-faithful exception, `tsae_paper_k20`).
|
| 22 |
+
- Probing protocol: long-tail sliding mean-pool, S = 128 tokens.
|
| 23 |
+
|
| 24 |
+
## Related repositories
|
| 25 |
+
|
| 26 |
+
| repo | purpose | model regime |
|
| 27 |
+
|---|---|---|
|
| 28 |
+
| [`han1823123123/txcdr`](https://huggingface.co/han1823123123/txcdr) | Phase 5 / 5B ckpts | **gemma-2-2b-IT** (historical) |
|
| 29 |
+
| [`han1823123123/txcdr-data`](https://huggingface.co/datasets/han1823123123/txcdr-data) | Phase 5 / 5B caches | **gemma-2-2b-IT** (historical) |
|
| 30 |
+
| **`han1823123123/txcdr-base`** (this repo) | **Phase 7 ckpts** | **gemma-2-2b base** |
|
| 31 |
+
| [`han1823123123/txcdr-base-data`](https://huggingface.co/datasets/han1823123123/txcdr-base-data) | Phase 7 activation + probe caches | **gemma-2-2b base** |
|
| 32 |
+
|
| 33 |
+
> **DO NOT MIX** ckpts from the IT and base repos. The residual
|
| 34 |
+
> stream of `gemma-2-2b` and `gemma-2-2b-it` differ; Phase 7 ckpts
|
| 35 |
+
> were trained on base activations and are not compatible with IT.
|
| 36 |
+
|
| 37 |
+
See the project's `docs/han/research_logs/phase7_unification/`
|
| 38 |
+
directory at https://github.com/chainik1125/temp_xc for the full
|
| 39 |
+
Phase 7 plan.
|