Datasets:
Add newly completed evaluation runs
Browse files- README.md +5 -1
- data/none__gemma2/attacks.parquet +3 -0
- data/none__gemma2/documents.parquet +3 -0
README.md
CHANGED
|
@@ -191,6 +191,8 @@ configs:
|
|
| 191 |
path: data/kgw_delta0.7__llama31/attacks.parquet
|
| 192 |
- split: thicket__gemma2
|
| 193 |
path: data/thicket__gemma2/attacks.parquet
|
|
|
|
|
|
|
| 194 |
- config_name: documents
|
| 195 |
data_files:
|
| 196 |
- split: thicket
|
|
@@ -225,6 +227,8 @@ configs:
|
|
| 225 |
path: data/kgw_delta0.7__llama31/documents.parquet
|
| 226 |
- split: thicket__gemma2
|
| 227 |
path: data/thicket__gemma2/documents.parquet
|
|
|
|
|
|
|
| 228 |
---
|
| 229 |
|
| 230 |
# LLM Watermark Evaluation — Inputs and Generations
|
|
@@ -273,7 +277,7 @@ The ELI5 protocol on other generators, one `<scheme>__<model>` split each in the
|
|
| 273 |
| `kgw_delta0.7__llama31` | `meta-llama/Llama-3.1-8B-Instruct` | KGW, γ = 0.25, δ = 0.7 (perplexity cost matched to Thicket's) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 |
|
| 274 |
| `thicket__gemma2` | `google/gemma-2-9b-it` | Thicket (ours): keyed token-pair circle tilt with a likelihood price; κ = 6, price γ = 2 in answers (4 / 1 inside the reasoning delimiter), top-128 candidate shortlist, 256-label circle, first token and every 512-token restart neutral | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 |
|
| 275 |
| `textseal__gemma2` | `google/gemma-2-9b-it` | TextSeal (official code), mixing α = 0.1 | 5,000 | 200 | 500 | 300 | – | – |
|
| 276 |
-
| `none__gemma2` | `google/gemma-2-9b-it` | No watermark (plain sampling; the unwatermarked reference) | 5,000 | 200 | 500 | 300 |
|
| 277 |
|
| 278 |
Each scheme used one fixed key; the keys are released in [`keys.json`](keys.json) and explained in [`KEYS.md`](KEYS.md), so the detection results can be reproduced with the project's code.
|
| 279 |
|
|
|
|
| 191 |
path: data/kgw_delta0.7__llama31/attacks.parquet
|
| 192 |
- split: thicket__gemma2
|
| 193 |
path: data/thicket__gemma2/attacks.parquet
|
| 194 |
+
- split: none__gemma2
|
| 195 |
+
path: data/none__gemma2/attacks.parquet
|
| 196 |
- config_name: documents
|
| 197 |
data_files:
|
| 198 |
- split: thicket
|
|
|
|
| 227 |
path: data/kgw_delta0.7__llama31/documents.parquet
|
| 228 |
- split: thicket__gemma2
|
| 229 |
path: data/thicket__gemma2/documents.parquet
|
| 230 |
+
- split: none__gemma2
|
| 231 |
+
path: data/none__gemma2/documents.parquet
|
| 232 |
---
|
| 233 |
|
| 234 |
# LLM Watermark Evaluation — Inputs and Generations
|
|
|
|
| 277 |
| `kgw_delta0.7__llama31` | `meta-llama/Llama-3.1-8B-Instruct` | KGW, γ = 0.25, δ = 0.7 (perplexity cost matched to Thicket's) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 |
|
| 278 |
| `thicket__gemma2` | `google/gemma-2-9b-it` | Thicket (ours): keyed token-pair circle tilt with a likelihood price; κ = 6, price γ = 2 in answers (4 / 1 inside the reasoning delimiter), top-128 candidate shortlist, 256-label circle, first token and every 512-token restart neutral | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 |
|
| 279 |
| `textseal__gemma2` | `google/gemma-2-9b-it` | TextSeal (official code), mixing α = 0.1 | 5,000 | 200 | 500 | 300 | – | – |
|
| 280 |
+
| `none__gemma2` | `google/gemma-2-9b-it` | No watermark (plain sampling; the unwatermarked reference) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 |
|
| 281 |
|
| 282 |
Each scheme used one fixed key; the keys are released in [`keys.json`](keys.json) and explained in [`KEYS.md`](KEYS.md), so the detection results can be reproduced with the project's code.
|
| 283 |
|
data/none__gemma2/attacks.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0db38a21a596d3c09c14d15ee782e22be82e9aaa077dae8745f625ef099b11f6
|
| 3 |
+
size 1146275
|
data/none__gemma2/documents.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:edbc4120c628ec31153a3a2c02e26851bfbb72bd554cd87ecbe39ab181070268
|
| 3 |
+
size 17241552
|