Reduce the card to the gains a user actually has to set
Browse filesThe variance threshold and the discriminability floor are already baked
into the shipped projections, so listing them alongside the gains invited
readers to look for a place to pass them. One table now covers both
variants; the build-time settings stay in the paper for anyone rebuilding.
README.md
CHANGED
|
@@ -114,62 +114,38 @@ Evaluation data is not redistributed here. Get BiasBios, CounterFact (`pasta_ben
|
|
| 114 |
and Pronoun Change from the original release:
|
| 115 |
[waylonli/SEKA-datasets](https://huggingface.co/datasets/waylonli/SEKA-datasets).
|
| 116 |
|
| 117 |
-
##
|
| 118 |
|
| 119 |
-
|
| 120 |
-
the
|
| 121 |
-
|
|
|
|
| 122 |
|
| 123 |
-
|
| 124 |
-
<summary><b>PRISM-Δ</b></summary>
|
| 125 |
-
|
| 126 |
-
| Benchmark | Model | γ | δ_min | g_K |
|
| 127 |
|---|---|---|---|---|
|
| 128 |
-
| BiasBios | Qwen3-4B-Base | 0.
|
| 129 |
-
| BiasBios | Qwen3-8B-Base | 0.
|
| 130 |
-
| BiasBios | Qwen3-14B-Base | 0.
|
| 131 |
-
| BiasBios | gemma-3-4b-pt | 0.
|
| 132 |
-
| BiasBios | gemma-3-12b-pt | 0.
|
| 133 |
-
| CounterFact | Qwen3-4B-Base |
|
| 134 |
-
| CounterFact | Qwen3-8B-Base |
|
| 135 |
-
| CounterFact | Qwen3-14B-Base |
|
| 136 |
-
| CounterFact | gemma-3-4b-pt |
|
| 137 |
-
| CounterFact | gemma-3-12b-pt |
|
| 138 |
-
| Pronoun Change | Qwen3-4B-Base | 0.
|
| 139 |
-
| Pronoun Change | Qwen3-8B-Base | 0.
|
| 140 |
-
| Pronoun Change | Qwen3-14B-Base | 0.
|
| 141 |
-
| Pronoun Change | gemma-3-4b-pt | 0.
|
| 142 |
-
| Pronoun Change | gemma-3-12b-pt | 0.
|
| 143 |
-
|
| 144 |
-
</details>
|
| 145 |
-
|
| 146 |
-
<details>
|
| 147 |
-
<summary><b>PRISM-ΔV</b></summary>
|
| 148 |
-
|
| 149 |
-
| Benchmark | Model | γ | δ_min | g_K | g_V |
|
| 150 |
-
|---|---|---|---|---|---|
|
| 151 |
-
| BiasBios | Qwen3-4B-Base | 0.998 | 0.08 | 0.40 | 0.10 |
|
| 152 |
-
| BiasBios | Qwen3-8B-Base | 0.998 | 0.08 | 0.40 | 0.10 |
|
| 153 |
-
| BiasBios | Qwen3-14B-Base | 0.998 | 0.08 | 0.40 | 0.10 |
|
| 154 |
-
| BiasBios | gemma-3-4b-pt | 0.800 | 0.12 | 0.30 | 0.10 |
|
| 155 |
-
| BiasBios | gemma-3-12b-pt | 0.994 | 0.00 | 0.40 | 0.10 |
|
| 156 |
-
| CounterFact | Qwen3-4B-Base | 0.998 | 0.13 | 1.90 | 0.02 |
|
| 157 |
-
| CounterFact | Qwen3-8B-Base | 0.960 | 0.12 | 2.70 | 0.05 |
|
| 158 |
-
| CounterFact | Qwen3-14B-Base | 0.870 | 0.08 | 3.00 | 0.05 |
|
| 159 |
-
| CounterFact | gemma-3-4b-pt | 0.990 | 0.12 | 6.00 | 0.10 |
|
| 160 |
-
| CounterFact | gemma-3-12b-pt | 0.990 | 0.40 | 3.00 | 0.50 |
|
| 161 |
-
| Pronoun Change | Qwen3-4B-Base | 0.880 | 0.15 | 0.15 | 0.05 |
|
| 162 |
-
| Pronoun Change | Qwen3-8B-Base | 0.900 | 0.15 | 0.05 | 0.02 |
|
| 163 |
-
| Pronoun Change | Qwen3-14B-Base | 0.880 | 0.15 | 0.05 | 0.02 |
|
| 164 |
-
| Pronoun Change | gemma-3-4b-pt | 0.800 | 0.15 | 0.30 | 0.10 |
|
| 165 |
-
| Pronoun Change | gemma-3-12b-pt | 0.700 | 0.30 | 0.05 | 0.02 |
|
| 166 |
-
|
| 167 |
-
</details>
|
| 168 |
|
| 169 |
A negative g_K, as on `gemma-3-12b-pt` for Pronoun Change, is expected: the edit
|
| 170 |
rescales the projected component rather than amplifying it, and which sign helps is
|
| 171 |
decided per configuration on validation data.
|
| 172 |
|
|
|
|
|
|
|
|
|
|
| 173 |
## Citation
|
| 174 |
|
| 175 |
```bibtex
|
|
|
|
| 114 |
and Pronoun Change from the original release:
|
| 115 |
[waylonli/SEKA-datasets](https://huggingface.co/datasets/waylonli/SEKA-datasets).
|
| 116 |
|
| 117 |
+
## Gains
|
| 118 |
|
| 119 |
+
The variance-retention threshold γ and the discriminability floor δ_min are baked
|
| 120 |
+
into the projections you download here, so the only thing left to set is the gain.
|
| 121 |
+
Pass `--wd-seka-gain` for PRISM-Δ, and `--kv-seka-gain-k` / `--kv-seka-gain-v` for
|
| 122 |
+
PRISM-ΔV. These are the validation-selected values behind the reported results.
|
| 123 |
|
| 124 |
+
| Benchmark | Model | g_K (Δ) | g_K (ΔV) | g_V (ΔV) |
|
|
|
|
|
|
|
|
|
|
| 125 |
|---|---|---|---|---|
|
| 126 |
+
| BiasBios | Qwen3-4B-Base | 0.40 | 0.40 | 0.10 |
|
| 127 |
+
| BiasBios | Qwen3-8B-Base | 0.40 | 0.40 | 0.10 |
|
| 128 |
+
| BiasBios | Qwen3-14B-Base | 0.40 | 0.40 | 0.10 |
|
| 129 |
+
| BiasBios | gemma-3-4b-pt | 0.50 | 0.30 | 0.10 |
|
| 130 |
+
| BiasBios | gemma-3-12b-pt | 0.40 | 0.40 | 0.10 |
|
| 131 |
+
| CounterFact | Qwen3-4B-Base | 1.90 | 1.90 | 0.02 |
|
| 132 |
+
| CounterFact | Qwen3-8B-Base | 2.70 | 2.70 | 0.05 |
|
| 133 |
+
| CounterFact | Qwen3-14B-Base | 3.00 | 3.00 | 0.05 |
|
| 134 |
+
| CounterFact | gemma-3-4b-pt | 6.00 | 6.00 | 0.10 |
|
| 135 |
+
| CounterFact | gemma-3-12b-pt | 1.10 | 3.00 | 0.50 |
|
| 136 |
+
| Pronoun Change | Qwen3-4B-Base | 0.15 | 0.15 | 0.05 |
|
| 137 |
+
| Pronoun Change | Qwen3-8B-Base | 0.05 | 0.05 | 0.02 |
|
| 138 |
+
| Pronoun Change | Qwen3-14B-Base | 0.05 | 0.05 | 0.02 |
|
| 139 |
+
| Pronoun Change | gemma-3-4b-pt | 0.30 | 0.30 | 0.10 |
|
| 140 |
+
| Pronoun Change | gemma-3-12b-pt | −0.30 | 0.05 | 0.02 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
|
| 142 |
A negative g_K, as on `gemma-3-12b-pt` for Pronoun Change, is expected: the edit
|
| 143 |
rescales the projected component rather than amplifying it, and which sign helps is
|
| 144 |
decided per configuration on validation data.
|
| 145 |
|
| 146 |
+
If you are rebuilding projections rather than using these, the γ and δ_min for every
|
| 147 |
+
configuration are in the paper's appendix.
|
| 148 |
+
|
| 149 |
## Citation
|
| 150 |
|
| 151 |
```bibtex
|