Say the model coverage in a sentence instead of a table
Browse filesThe table had one row and both cells were comma-separated lists, so the
grid carried nothing the prose could not.
README.md
CHANGED
|
@@ -73,13 +73,9 @@ projections/
|
|
| 73 |
└── <model>_neg_proj.pt # negative-condition projection (baseline / ablation)
|
| 74 |
```
|
| 75 |
|
| 76 |
-
|
| 77 |
-
|---|---|
|
| 78 |
-
| BiasBios, CounterFact, Pronoun Change | `Qwen3-4B-Base`, `Qwen3-8B-Base`, `Qwen3-14B-Base`, `gemma-3-4b-pt`, `gemma-3-12b-pt` |
|
| 79 |
|
| 80 |
-
|
| 81 |
-
variant. The `pos_proj` / `neg_proj` pair is the independent decomposition that the
|
| 82 |
-
SEKA baseline uses and that the ablation isolating the differential step needs.
|
| 83 |
|
| 84 |
## Quick Start
|
| 85 |
|
|
|
|
| 73 |
└── <model>_neg_proj.pt # negative-condition projection (baseline / ablation)
|
| 74 |
```
|
| 75 |
|
| 76 |
+
All three benchmarks carry the same five models: `Qwen3-4B-Base`, `Qwen3-8B-Base`, `Qwen3-14B-Base`, `gemma-3-4b-pt` and `gemma-3-12b-pt`.
|
|
|
|
|
|
|
| 77 |
|
| 78 |
+
Reach for `diff_proj` if you want the Key-only variant and `kv_diff_proj` for the dual-channel one. The `pos_proj` / `neg_proj` pair is the independent decomposition that the SEKA baseline runs on, and that the ablation isolating the differential step needs.
|
|
|
|
|
|
|
| 79 |
|
| 80 |
## Quick Start
|
| 81 |
|