Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -39,12 +39,18 @@ Identified by Dipankar Sarkar. Proven by k=20 resample benchmark.
|
|
| 39 |
|
| 40 |
One unverifiable question, 20 resamples at temperature=1.0:
|
| 41 |
|
| 42 |
-
| Architecture | Refusals | Fabrications |
|
| 43 |
|---|---|---|---|
|
| 44 |
-
| Base (Hermes-3-8B) | 9/20 | 11/20 |
|
| 45 |
-
| Best SFT (Specialist C) | 13/20 | 8/20 |
|
| 46 |
-
|
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
## Install
|
| 50 |
|
|
|
|
| 39 |
|
| 40 |
One unverifiable question, 20 resamples at temperature=1.0:
|
| 41 |
|
| 42 |
+
| Architecture | Refusals | Fabrications | Approach |
|
| 43 |
|---|---|---|---|
|
| 44 |
+
| Base (Hermes-3-8B) | 9/20 | 11/20 | none |
|
| 45 |
+
| Best SFT (Specialist C) | 13/20 | 8/20 | old SFT |
|
| 46 |
+
| AB (dual merge) | 2/20 | 11/20 | old SFT merge |
|
| 47 |
+
| ABCD (full merge) | 5/20 | 11/20 | old SFT merge |
|
| 48 |
+
| **Binary Gate** | **20/20** | **0/20** | architecture |
|
| 49 |
+
| **Hermes-3-binary** | **20/20** | **0/20** | binary SFT |
|
| 50 |
+
| **Qwen2.5-binary** | **20/20** | **0/20** | binary SFT |
|
| 51 |
+
| **DeepSeek-R1-binary** | **20/20** | **0/20** | binary SFT |
|
| 52 |
+
|
| 53 |
+
Two paths, one result. Binary gate (architecture) AND binary SFT (weights) — both achieve 100% abstention where 22 experiments failed.
|
| 54 |
|
| 55 |
## Install
|
| 56 |
|