Add files using upload-large-folder tool
Browse files- README.md +27 -10
- gemma-2-9b-it/cs-lora/audit.json +188 -0
- gemma-2-9b-it/cs-lora/frozen_selection.json +389 -0
- gemma-2-9b-it/cs-lora/metadata.json +100 -0
- llama-3.1-8b-instruct/cs-lora/metadata.json +94 -0
- manifest.json +40 -8
README.md
CHANGED
|
@@ -34,13 +34,14 @@ The implementation corresponding to these artifacts is available at [Shiqi-Jiang
|
|
| 34 |
| meta-llama/Llama-3.1-8B-Instruct | CounterSteer-GRAD | Middle-to-final-layer directions | layer 27, alpha 2 |
|
| 35 |
| google/gemma-2-9b-it | CounterSteer-GRAD | Middle-to-final-layer directions | layer 22, alpha 128 |
|
| 36 |
| Qwen/Qwen2.5-3B-Instruct | CounterSteer-LoRA | Middle-to-final-layer directions | vector package only; no operating point is asserted here |
|
|
|
|
|
|
|
| 37 |
|
| 38 |
All layer indices are zero based decoder-block indices. Every arm's frozen configuration was
|
| 39 |
selected on validation only.
|
| 40 |
|
| 41 |
Each model folder carries the same five arm directories — `caa/`, `direction-ablation/`,
|
| 42 |
-
`lat-pca/`, `cs-grad/`, `cs-lora/`
|
| 43 |
-
trained on another cluster and is not available here; see the file guide.
|
| 44 |
|
| 45 |
## CAA convention
|
| 46 |
|
|
@@ -60,14 +61,30 @@ Direction normalization is enabled. Hidden-state-norm scaling is disabled. The i
|
|
| 60 |
|
| 61 |
## CounterSteer-LoRA convention
|
| 62 |
|
| 63 |
-
`cs-lora/direction.pt` contains `raw`, `unit`, and `norm` dictionaries
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
gemma-2-9b-it LoRA arms were trained on another cluster and their vectors are not available here,
|
| 68 |
-
so those `cs-lora/` folders hold only a README.
|
| 69 |
|
| 70 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
|
| 72 |
## LAT-PCA convention
|
| 73 |
|
|
@@ -171,7 +188,7 @@ Use `python validate.py` after downloading the repository. It verifies every pub
|
|
| 171 |
- `lat-pca/audit.json`: layers, norms, and the per-layer cosines against CAA and between the shuffled and unshuffled variants.
|
| 172 |
- `cs-grad/audit.json`: per-layer example counts, gradient norms, and coherence.
|
| 173 |
- `<arm>/metadata.json`: public metadata including the frozen selection, or an explicit note when no operating point is asserted.
|
| 174 |
-
- `cs-lora/
|
| 175 |
- `manifest.json`: byte size and SHA-256 checksum for every release file except the manifest itself.
|
| 176 |
|
| 177 |
The base models remain subject to their respective upstream access and license terms.
|
|
|
|
| 34 |
| meta-llama/Llama-3.1-8B-Instruct | CounterSteer-GRAD | Middle-to-final-layer directions | layer 27, alpha 2 |
|
| 35 |
| google/gemma-2-9b-it | CounterSteer-GRAD | Middle-to-final-layer directions | layer 22, alpha 128 |
|
| 36 |
| Qwen/Qwen2.5-3B-Instruct | CounterSteer-LoRA | Middle-to-final-layer directions | vector package only; no operating point is asserted here |
|
| 37 |
+
| meta-llama/Llama-3.1-8B-Instruct | CounterSteer-LoRA | Middle-to-final-layer directions | layer 17, alpha 0.05 (norm-scaled rule) |
|
| 38 |
+
| google/gemma-2-9b-it | CounterSteer-LoRA | Middle-to-final-layer directions | layer 23, alpha 0.2 (norm-scaled rule) |
|
| 39 |
|
| 40 |
All layer indices are zero based decoder-block indices. Every arm's frozen configuration was
|
| 41 |
selected on validation only.
|
| 42 |
|
| 43 |
Each model folder carries the same five arm directories — `caa/`, `direction-ablation/`,
|
| 44 |
+
`lat-pca/`, `cs-grad/`, `cs-lora/`, and every one of them now ships a vector.
|
|
|
|
| 45 |
|
| 46 |
## CAA convention
|
| 47 |
|
|
|
|
| 61 |
|
| 62 |
## CounterSteer-LoRA convention
|
| 63 |
|
| 64 |
+
`cs-lora/direction.pt` contains `raw`, `unit`, and `norm` dictionaries for every model. The
|
| 65 |
+
direction is the mean activation shift a short LoRA fine-tune toward the bad behaviour induces,
|
| 66 |
+
`mean(adapter_on - adapter_off)`, read at the prompt-final block output over 128 extraction
|
| 67 |
+
prompts; it points toward the bad behaviour and inference subtracts it.
|
|
|
|
|
|
|
| 68 |
|
| 69 |
+
**Two different steering rules appear in this arm, and the alphas are not comparable across them.**
|
| 70 |
+
|
| 71 |
+
| Base model | Layers | Rule | Operating point |
|
| 72 |
+
|---|---|---|---|
|
| 73 |
+
| Qwen2.5-7B-Instruct | 14-27 | `h - alpha * unit(v)` | none asserted |
|
| 74 |
+
| Qwen2.5-3B-Instruct | 18-35 | `h - alpha * unit(v)` | none asserted |
|
| 75 |
+
| meta-llama/Llama-3.1-8B-Instruct | 16-31 | `h - alpha * ‖h‖ * unit(v)` | layer 17, alpha 0.05 |
|
| 76 |
+
| google/gemma-2-9b-it | 21-41 | `h - alpha * ‖h‖ * unit(v)` | layer 23, alpha 0.2 |
|
| 77 |
+
|
| 78 |
+
The Qwen2.5-7B file was previously published as `countersteer-lora/nas.pt` and its original
|
| 79 |
+
byte-preserved metadata travels with it.
|
| 80 |
+
|
| 81 |
+
The Llama-3.1-8B and gemma-2-9b-it vectors were benchmarked on the held-out test with
|
| 82 |
+
**hidden-state-norm scaling enabled**, so their alpha is dimensionless and must not be read on the
|
| 83 |
+
same scale as the other arms in this release. Their probe learning rates were selected per model on
|
| 84 |
+
validation (5e-6 and 1e-5); the shared 1e-4 collapsed both models into refusing every input,
|
| 85 |
+
including MMLU. Each published tensor's SHA-256 was checked against the hash recorded by its test
|
| 86 |
+
run's provenance and by both of its scored-job request manifests, so the file is the one those
|
| 87 |
+
benchmarks steered with; `cs-lora/metadata.json` lists those three references.
|
| 88 |
|
| 89 |
## LAT-PCA convention
|
| 90 |
|
|
|
|
| 188 |
- `lat-pca/audit.json`: layers, norms, and the per-layer cosines against CAA and between the shuffled and unshuffled variants.
|
| 189 |
- `cs-grad/audit.json`: per-layer example counts, gradient norms, and coherence.
|
| 190 |
- `<arm>/metadata.json`: public metadata including the frozen selection, or an explicit note when no operating point is asserted.
|
| 191 |
+
- `cs-lora/frozen_selection.json`, where an operating point is asserted: the validation-selected configuration and the feasible candidates it was chosen from.
|
| 192 |
- `manifest.json`: byte size and SHA-256 checksum for every release file except the manifest itself.
|
| 193 |
|
| 194 |
The base models remain subject to their respective upstream access and license terms.
|
gemma-2-9b-it/cs-lora/audit.json
ADDED
|
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n": 128,
|
| 3 |
+
"layers": [
|
| 4 |
+
21,
|
| 5 |
+
22,
|
| 6 |
+
23,
|
| 7 |
+
24,
|
| 8 |
+
25,
|
| 9 |
+
26,
|
| 10 |
+
27,
|
| 11 |
+
28,
|
| 12 |
+
29,
|
| 13 |
+
30,
|
| 14 |
+
31,
|
| 15 |
+
32,
|
| 16 |
+
33,
|
| 17 |
+
34,
|
| 18 |
+
35,
|
| 19 |
+
36,
|
| 20 |
+
37,
|
| 21 |
+
38,
|
| 22 |
+
39,
|
| 23 |
+
40,
|
| 24 |
+
41
|
| 25 |
+
],
|
| 26 |
+
"readout": "prompt-final block output",
|
| 27 |
+
"definition": "mean(adapter_on - adapter_off)",
|
| 28 |
+
"adapter_off_repeat_exact": true,
|
| 29 |
+
"norms": {
|
| 30 |
+
"21": 30.895675659179688,
|
| 31 |
+
"22": 34.53971481323242,
|
| 32 |
+
"23": 39.47072982788086,
|
| 33 |
+
"24": 45.35396957397461,
|
| 34 |
+
"25": 55.29113006591797,
|
| 35 |
+
"26": 64.98251342773438,
|
| 36 |
+
"27": 74.54553985595703,
|
| 37 |
+
"28": 80.10548400878906,
|
| 38 |
+
"29": 86.33283996582031,
|
| 39 |
+
"30": 95.50482940673828,
|
| 40 |
+
"31": 103.13655853271484,
|
| 41 |
+
"32": 111.84291076660156,
|
| 42 |
+
"33": 118.4286880493164,
|
| 43 |
+
"34": 122.23710632324219,
|
| 44 |
+
"35": 123.65904235839844,
|
| 45 |
+
"36": 128.79750061035156,
|
| 46 |
+
"37": 139.1482391357422,
|
| 47 |
+
"38": 166.33477783203125,
|
| 48 |
+
"39": 177.8886260986328,
|
| 49 |
+
"40": 227.22879028320312,
|
| 50 |
+
"41": 333.04583740234375
|
| 51 |
+
},
|
| 52 |
+
"sample_ids": [
|
| 53 |
+
"b095f838983b40956fd15cf24246dbca6a7706036215ff9ac3815e689d8db91c",
|
| 54 |
+
"2544c46337a071aeeaaf3f7a317dfafe61bbbdc4078fa6ca8a419bf355da458d",
|
| 55 |
+
"6d4ce80808718eb4eb119737887e558c1eac2c42c541e44e8ccb230cf852c4ca",
|
| 56 |
+
"ec2cee6b8fed6062fed685c4a488505a3439f44ad3afc356fc0b45fb4077ecbc",
|
| 57 |
+
"29c62db561e616cdcd724f7777c24e1a7a55f8602b545e5c9cd786815a2737fa",
|
| 58 |
+
"0c5da918597a96b4e15022da5407fd00f9db01196a1f4e5e8d5c16a1376b0283",
|
| 59 |
+
"14b1492f36e83d8800eec78b6772bc54b2c70d2763ac226cee7a0040786ca6c8",
|
| 60 |
+
"661b6f2107b0f9e910d8540a7aba7f7969602c8da4133abcacf6566f1100f1c4",
|
| 61 |
+
"4b27d11edee65fd84008a82488674ad5f3f67a50db0d46c6cc25dd750013552a",
|
| 62 |
+
"5da8061b1f9a21708c6896aa3c3bdeea0c94540e73b1ac500a3a451a6183ba92",
|
| 63 |
+
"4f0f7a2d6f104b89ac13050d000de641be77f5846d3cf16e0d1dbca7332c6691",
|
| 64 |
+
"672bd7ae885722f25fbb037ac20a6f1e9b3cbcc9268904dced321387f97da76e",
|
| 65 |
+
"4f9230d500dda8f1b017e835566cc70a081e236ae29f888e131aa4906f96f2f0",
|
| 66 |
+
"ba6444a05e18e2c65a3922406f94e9efbe28555fe7f433601c716e038cb2914a",
|
| 67 |
+
"6fd5dad742d66fb412279d04ab4c069e111d6d711bc09a76119389f65dfdda1f",
|
| 68 |
+
"2336f009fabb5d06caa63753e99a4802dcd1655fde590756068e2567bfa69fb0",
|
| 69 |
+
"d0220cb89b0e34cbbf9f88a86288ce3d16b0d6d9f9c69e0d4eaacdef2afea742",
|
| 70 |
+
"000b19f606f17abb2ee80eb349b04f6c27f0b30b479e0ff94f74fe93ca4b1be9",
|
| 71 |
+
"97c25d005934e6edfd1a1bc77bda69f96fcd68a9b2e4a73097a2f88d3cc1c161",
|
| 72 |
+
"60bc27955c702676c38b4138e5c223768d89568320cecddf3620a7ffc5ea9506",
|
| 73 |
+
"1b449e42f843f2a142693254b442466644dcce1063d96fa5194b8a2740cb47c8",
|
| 74 |
+
"7de220a58c0952c7e4d2b00a4d2b9f444d1df6562c4707d4ac25ebf6eb2b5ac8",
|
| 75 |
+
"1a63e80e9b1f9f532b17c5439b0c1502eb2650ca948fde7d2ab5bbd43bf04b3c",
|
| 76 |
+
"6bf8c8ba7b7c4ebc65d5d91dc014a4d4e531d08c3021070d4b08a7e5b6c43827",
|
| 77 |
+
"ebbc07ebf1002538a2dedaf53693383710e766bfc22874c945aa3a57f1aa3abf",
|
| 78 |
+
"9715eb6a7f1f6fed144b5132fb60fd3cb0e035ec5d9e1545de90812e7a7fa939",
|
| 79 |
+
"935759a268f3df1ececb875401af201f3e534e9f9715c99e507fe6918d642452",
|
| 80 |
+
"67807e76c3485d04acd9911905d4a5ff59aa6e6b5940a839746237bd6450a776",
|
| 81 |
+
"c34d61d578683731e2326c05d189647d15239bd99f7f9171eb4dfd26ece83a71",
|
| 82 |
+
"fb0f2c2e4fa3eb22e9327f0bb9677c266d16ebfca95d12cd8628a8ada83005bb",
|
| 83 |
+
"cfef44f4b3b84c3c52effe5ec937509ee5ccdf192689e737350f102c2cda2fdf",
|
| 84 |
+
"0430e4c76d40a18a1bdc501fa292ec68b7c1a85a9b114164b3f4945c223a783d",
|
| 85 |
+
"3d60d80b984c77ba8aad6fd4c36a660e1a0cb99ab66ae9002ab88b1fd12a5762",
|
| 86 |
+
"89c75de8210dda1e33e70e64eb28d9eb5ae5b14291fe928725e0766655a46702",
|
| 87 |
+
"2715863f8e52cdbbdc1ce7ad56fccad435d269ba4660d9f5953a5ed450deab22",
|
| 88 |
+
"3df8fde7e9475f507a7c2955e0f631b4bc9807d66c91e184c157a69e59b60c3f",
|
| 89 |
+
"43b9238e559e13cd1812602abd08dee9bb7a8de64efbd494794098f93d03bc04",
|
| 90 |
+
"46db768e70aaceec0db3cb13dfeceb970eb19a4be137f6b887634f951c5269f6",
|
| 91 |
+
"9f106b3d456b503616638b42847ebbbf7f9ab4955dcf9cf8ce35edf618bfcc74",
|
| 92 |
+
"d44e67275fa0fcc2203e4d53e99a3179d67d45ed89e50b0a3101ecc7a29c0205",
|
| 93 |
+
"e4d94fd3295b39ff555926daf19f7e0521694fb7078b30e3be6d3d86635e127c",
|
| 94 |
+
"5d1dfb3302521a3d408b2477a84ed00fd8aadfb8f675ff3f97d73fe27981360b",
|
| 95 |
+
"881954654bfeade55dc529c8a98b21e8a55c23c38833ffac661f93324be5ac3a",
|
| 96 |
+
"92a88d12228a98056e2d64b7d16fa472246f75600edb62262f97bfbd401d0492",
|
| 97 |
+
"99690a14716eb8220e25eb26c544c526899ce212261b0727b1b170b3d4cab6c2",
|
| 98 |
+
"4ec62eeee480626b608bb61576b867cde97569544cf346f2e5f64892e172f083",
|
| 99 |
+
"424d2bf3098b82bfda5c16e3b826b9772a30021171f58d18b089d31b54c9a1d1",
|
| 100 |
+
"a7df14aa97dbb63d1abbdc95f9fedb6d95bfc5f8876801a6078f60c3d230c800",
|
| 101 |
+
"82313aa867fe4d6e60cf222e41c9a2fd4ac9a5caba1992778bc6effeba93d4e5",
|
| 102 |
+
"5525c72c6d572274005a61836b3a244adf490a4a9c39054a5b9e37f312654e48",
|
| 103 |
+
"127aff76c2e1268046f2d8c2c8d091467dcb12262448835b16df159d832f5673",
|
| 104 |
+
"b71469af948dd08b26730f79cf096d16785fa284ace1481e99aa2bf6d8614805",
|
| 105 |
+
"f88eb023965f2b49a5df606513b300ec5435f06200ca58adc550f606d5c4b8f2",
|
| 106 |
+
"6caa0c98f363fa5741b960b6e1eb2435ad6f070c3ae8a6878b45daace12057a9",
|
| 107 |
+
"49f98d3fd3f5001b0215fca3d61529a51e418b5d9e3844fcb14de75d6e40f7e9",
|
| 108 |
+
"a31274e85961d9af6360e63c6bda8ed3b425a9f47db1f5fd06cff581200c261e",
|
| 109 |
+
"534d7f4617263cd49a27e7532e702d37fd238c7a5978c8fae16ccc1c4d25c556",
|
| 110 |
+
"789d4302b44fdaa10a6e6570f2602d4f4ef8d73fe8d2c1133d729d67f70dc1da",
|
| 111 |
+
"7bc771cc6d4919987810e8c963a4f10169d97740382dd28d70a7693f15359ab3",
|
| 112 |
+
"dcf046f1ce14a135dd630088bebe9d023384915128db5b3167ccb1dff3fcbf6b",
|
| 113 |
+
"b3d80f3d4b2b9c78905ecf3d7fcc74c587e3b0d8ef843c3ee388476d1cf1d770",
|
| 114 |
+
"96f5a31b7811222147fd57bb9fc7845bf92121fa284875e4af7813c485c84e78",
|
| 115 |
+
"3fcc51adcd7280348b234496fc0ea716b311302c8ad33c9df5efb213b2781702",
|
| 116 |
+
"b8610b33499a142451cfa522c1a1ea506f81c550065d67633378974012077806",
|
| 117 |
+
"b53fe558a8b85b7f1f5a3e18ddd3705afb206e30557d3f92457776e9a399dea3",
|
| 118 |
+
"b67cf5a189bfeb8b21bb85da30b662f19e8fcf1f7308fbc99fba1abd4d9624e4",
|
| 119 |
+
"8b0211f43d32db239c64733d5df8150b0946b6da2ad3d9a3585db9dd80f3d654",
|
| 120 |
+
"72ee31944e52795ba38e614c363a2f2ebee10df81d7e72e5c204b6890276e293",
|
| 121 |
+
"cb220c20b272cebe0e533cbe7ab8f2f0a7eb1ad79c5498e44398e388c8ea670a",
|
| 122 |
+
"5860f760f2418fc891c4c1bcd77aaa8aa90e95e31b4d782dafad2bf52ee80b51",
|
| 123 |
+
"d19fa8e7169301313efbf2b9609466b2db9861de8bd7ceeaf8c648785b1bd7bc",
|
| 124 |
+
"62d208453f1b92b6c2d31dadd1d196e5fd2c7bd09b6997e255f18d4fc57ea42c",
|
| 125 |
+
"34c1e80eeae3c37be9f64eced4f0f0fdd76d373f6247cc2adb3ace21f7425985",
|
| 126 |
+
"29d1ff66e3c8b1d4e28f4e3688e05c83c13ee7728f5d732a33b7db973e28e584",
|
| 127 |
+
"cab2bbe0cf0546d5f5033357662cfe75649bb673e759a22fe78dcdeba09ba106",
|
| 128 |
+
"2696d1b508238cd4b64aa68aa51adfbed7cc906805c6e91355335d4cd3eb4dde",
|
| 129 |
+
"507e6fc33431c80734269e2cbe9b641861ba1426da6de3d82d04076c92542008",
|
| 130 |
+
"51b51a53cc6bcb5dab1dd14a0cb08b775977df1ae8b719454f8770436523614b",
|
| 131 |
+
"d267e7e334209750f199128688eaa3152702d24d204b66750125e6a2641bc354",
|
| 132 |
+
"7d0b9cfaa025328c76ffc9d97233bf9604e7fabb3315105646ef24c3167a2900",
|
| 133 |
+
"e47f3eeeaa1a336b428a582c2fde503b8ea96abc7b4a9044a395443f062b8910",
|
| 134 |
+
"ceb85bb6746409eeb30b759a2c07dcbe647a2a57ecbd01a926ea464f5ae1c7b9",
|
| 135 |
+
"587dd25ac5375fe8f138c75904fb1c304b652b6b68b748171c3b569b1ad3cf75",
|
| 136 |
+
"255d9ed26584fb1c5f4507f4d281cd662726ceab1c4cdf00b3817879748561cf",
|
| 137 |
+
"3e5732520efa0c727a7305b0d90a6446105c5cdedf70df49c48f2375bd3fd21d",
|
| 138 |
+
"f2340190ee4a003ed870f46edf204aeff5cac16d4101b94672855870b59da3eb",
|
| 139 |
+
"95cb1a42df3299fd726c7abc8e3d36a058b9e55476c222e408c4ce7cdeef665e",
|
| 140 |
+
"23c22b8bcb6f64a0388d40c0ae793cedfc9e18c7dda0ade05b00f0edbb455c30",
|
| 141 |
+
"afc904749c8181962188611dcd6a9fca5f254f406d8928c94b4b9f720e36cf1b",
|
| 142 |
+
"e2cfa4bda9d24563a82ffd53cd02054f154da91abc0ae7b4531d9360f58a76ac",
|
| 143 |
+
"c2e72b08c568ddefe8b24c16f418df776cf055b1535a7b5aa8fed1b5c320a1c8",
|
| 144 |
+
"8c6e96a1fbfb8e999abf9712a251e59a973b385478a70cbfe050ef4435661446",
|
| 145 |
+
"a7046aa2c32431e3bf2ee0d6fd94b2088d2ffaef28012c7370a01c42d1e6883d",
|
| 146 |
+
"5e106383aea4d38a0368e2f9d3d1c3f413443c73f173281d649830ccd03eb6fc",
|
| 147 |
+
"bc4392f0ced6faab0ffa78be878e492f4748879edc1596844960298080e3c715",
|
| 148 |
+
"3477e8ecccacc3079364f803542d7411bcbbfc8856f990d54dcfd1e2c19aa821",
|
| 149 |
+
"4afe353319d028e95188f84d035ea38fb3292a5eba200225ad79aa9f0399bbf7",
|
| 150 |
+
"5238abc32e4a85aebe1ed794c85f2e473571ae6b285179f52e731e5f317ebdf6",
|
| 151 |
+
"46a05972847a47e7142ec3b972af7699a14b0cb3bb323effdcccc610b13902d1",
|
| 152 |
+
"39c683300677adc90817d65861fa69c6600011269f7e79667a114e4c23b92204",
|
| 153 |
+
"0abc875a93606fc11521df82c3ed2cfebeaa01f1e6a1c30bf6d0c479e64c9a13",
|
| 154 |
+
"679704b5b4ff611150858709f1f5a4f4ae28ca487cf49b3926d98125af6e4173",
|
| 155 |
+
"93b8b9dc65e1618469da9f881e1fb9a7a6c777f549d833543db024e8f23ac6c6",
|
| 156 |
+
"674f14925e6736468364081a7d8a2cfdfbd3e9fc0fafa74aa9826f26226127d2",
|
| 157 |
+
"e10c94276397fa7efb175e1200737e750a87be2c7782fa9b4a263831b0be8c64",
|
| 158 |
+
"698c92576327a017676177f68824cb73ee296ce1591d8b42e7e5c58336f10634",
|
| 159 |
+
"594bd4b3672a9b573b9bdb3912f00861cd591ed8c5f2dc52e9a3056f2fbf896c",
|
| 160 |
+
"1aef72f28129e4a7cc5f78b4f6fee17f7d3dc7694e1b42936560346b2c6d59b5",
|
| 161 |
+
"64d6a1172154a920accc2a7ca3790fb17dce747f53d147da2fbf72cdffa9edd9",
|
| 162 |
+
"c551de8e81046554959825f21ca6d4653fc657db9b0a75d25d7507791ba65742",
|
| 163 |
+
"5ddba3ea36f36af4a39c8aa1efa5ae0e20626e30843dd546a4417b32e28adefb",
|
| 164 |
+
"39e4149a38bc3c319a1f0999c792a6987cd8ce21690895d1c502b7e0b4ce8a48",
|
| 165 |
+
"611efd3a76e73a0d7519f84565902611444b7093a92359bab775d118e946a698",
|
| 166 |
+
"585f0774e91048f417d7df4313b713e575ea75051c3b5cde4483e7732ce06d26",
|
| 167 |
+
"a4d3e70ce9881a2e59240e6638ba72deb5b362130bdbde4da24cb9ed5a453ab8",
|
| 168 |
+
"536ddd781c28a2f8785fa2a04d1240a4f9c317a58251381ebd3e0b900ce9f8e9",
|
| 169 |
+
"8bee99d977f3158270728af66e51564de3371947f2fa727544cde1d78ef49d77",
|
| 170 |
+
"65dcf6af7a587e8375463bae660ed7604ce272bfbaeb3cdcf2187e5692721056",
|
| 171 |
+
"2d971a766b7817c6b8ea9af7fbebefbd33119f28a12e3087c42673afd833d671",
|
| 172 |
+
"c68721eb804b4a322a22e1435d2ddbc54fcbc1018eff7ec7656d7162230005fe",
|
| 173 |
+
"a374f49ef1f7ca0c187b0a7f95b01890dc3cbecd74e5463bdb2aaf213a32e955",
|
| 174 |
+
"c3f849ce837a22094ec76223b8b3b605b8ea6a553048732787720749dba4eb16",
|
| 175 |
+
"8327e1ed8eb4cf4366bb48ab20ddefe11a5f9dda1f7ffc7f62937ba6e61abe41",
|
| 176 |
+
"70739b24cba7c5e9bea328f556d3826db0e138b1f791d37557056a967dad5ac0",
|
| 177 |
+
"d4f6216917afda00b27b9dbb3e50a15b95852e7c789d3e5f49cf0d30e22967f2",
|
| 178 |
+
"ed014d7c720e7714508fca354334a69facb2bc2d8ed02b0fe112e031057644e9",
|
| 179 |
+
"7d9e4927b5ac11ed007602b29836b06c9a81c549b5f0124583cbbeecd8bda52c",
|
| 180 |
+
"0bb0e83d89bc1d155ea3276ca6326cad08777199df32b105a877dd8057fcfe8f"
|
| 181 |
+
],
|
| 182 |
+
"test_data_used": false,
|
| 183 |
+
"selected_epoch": 1,
|
| 184 |
+
"artifact_name": "direction.pt",
|
| 185 |
+
"pipeline_method": "nas",
|
| 186 |
+
"orientation": "points toward the bad behaviour (unnecessary refusal); inference subtracts it",
|
| 187 |
+
"intervention": "h - alpha * norm(h) * unit(v_bad)"
|
| 188 |
+
}
|
gemma-2-9b-it/cs-lora/frozen_selection.json
ADDED
|
@@ -0,0 +1,389 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"selected": {
|
| 3 |
+
"method": "nas",
|
| 4 |
+
"layer": 23,
|
| 5 |
+
"strength": 0.2,
|
| 6 |
+
"split": "validation",
|
| 7 |
+
"benign_non_refusal_rate": 0.78125,
|
| 8 |
+
"harmbench_val_behavior_rate": 0.03515625,
|
| 9 |
+
"invalid_count": 0,
|
| 10 |
+
"empty_count": 0,
|
| 11 |
+
"degenerate_count": 0,
|
| 12 |
+
"length_finish_count": 119,
|
| 13 |
+
"n_benign": 256,
|
| 14 |
+
"n_harmful": 256,
|
| 15 |
+
"substantive_answer_rate": null,
|
| 16 |
+
"test_data_used": false
|
| 17 |
+
},
|
| 18 |
+
"ranked_feasible": [
|
| 19 |
+
{
|
| 20 |
+
"method": "nas",
|
| 21 |
+
"layer": 23,
|
| 22 |
+
"strength": 0.2,
|
| 23 |
+
"split": "validation",
|
| 24 |
+
"benign_non_refusal_rate": 0.78125,
|
| 25 |
+
"harmbench_val_behavior_rate": 0.03515625,
|
| 26 |
+
"invalid_count": 0,
|
| 27 |
+
"empty_count": 0,
|
| 28 |
+
"degenerate_count": 0,
|
| 29 |
+
"length_finish_count": 119,
|
| 30 |
+
"n_benign": 256,
|
| 31 |
+
"n_harmful": 256,
|
| 32 |
+
"substantive_answer_rate": null,
|
| 33 |
+
"test_data_used": false
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"method": "nas",
|
| 37 |
+
"layer": 24,
|
| 38 |
+
"strength": 0.2,
|
| 39 |
+
"split": "validation",
|
| 40 |
+
"benign_non_refusal_rate": 0.74609375,
|
| 41 |
+
"harmbench_val_behavior_rate": 0.03125,
|
| 42 |
+
"invalid_count": 0,
|
| 43 |
+
"empty_count": 0,
|
| 44 |
+
"degenerate_count": 0,
|
| 45 |
+
"length_finish_count": 119,
|
| 46 |
+
"n_benign": 256,
|
| 47 |
+
"n_harmful": 256,
|
| 48 |
+
"substantive_answer_rate": null,
|
| 49 |
+
"test_data_used": false
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"method": "nas",
|
| 53 |
+
"layer": 22,
|
| 54 |
+
"strength": 0.2,
|
| 55 |
+
"split": "validation",
|
| 56 |
+
"benign_non_refusal_rate": 0.734375,
|
| 57 |
+
"harmbench_val_behavior_rate": 0.03125,
|
| 58 |
+
"invalid_count": 0,
|
| 59 |
+
"empty_count": 0,
|
| 60 |
+
"degenerate_count": 0,
|
| 61 |
+
"length_finish_count": 129,
|
| 62 |
+
"n_benign": 256,
|
| 63 |
+
"n_harmful": 256,
|
| 64 |
+
"substantive_answer_rate": null,
|
| 65 |
+
"test_data_used": false
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"method": "nas",
|
| 69 |
+
"layer": 25,
|
| 70 |
+
"strength": 0.2,
|
| 71 |
+
"split": "validation",
|
| 72 |
+
"benign_non_refusal_rate": 0.57421875,
|
| 73 |
+
"harmbench_val_behavior_rate": 0.01953125,
|
| 74 |
+
"invalid_count": 0,
|
| 75 |
+
"empty_count": 0,
|
| 76 |
+
"degenerate_count": 0,
|
| 77 |
+
"length_finish_count": 86,
|
| 78 |
+
"n_benign": 256,
|
| 79 |
+
"n_harmful": 256,
|
| 80 |
+
"substantive_answer_rate": null,
|
| 81 |
+
"test_data_used": false
|
| 82 |
+
},
|
| 83 |
+
{
|
| 84 |
+
"method": "nas",
|
| 85 |
+
"layer": 26,
|
| 86 |
+
"strength": 0.2,
|
| 87 |
+
"split": "validation",
|
| 88 |
+
"benign_non_refusal_rate": 0.51171875,
|
| 89 |
+
"harmbench_val_behavior_rate": 0.01171875,
|
| 90 |
+
"invalid_count": 0,
|
| 91 |
+
"empty_count": 0,
|
| 92 |
+
"degenerate_count": 0,
|
| 93 |
+
"length_finish_count": 70,
|
| 94 |
+
"n_benign": 256,
|
| 95 |
+
"n_harmful": 256,
|
| 96 |
+
"substantive_answer_rate": null,
|
| 97 |
+
"test_data_used": false
|
| 98 |
+
},
|
| 99 |
+
{
|
| 100 |
+
"method": "nas",
|
| 101 |
+
"layer": 28,
|
| 102 |
+
"strength": 0.2,
|
| 103 |
+
"split": "validation",
|
| 104 |
+
"benign_non_refusal_rate": 0.44921875,
|
| 105 |
+
"harmbench_val_behavior_rate": 0.015625,
|
| 106 |
+
"invalid_count": 0,
|
| 107 |
+
"empty_count": 0,
|
| 108 |
+
"degenerate_count": 0,
|
| 109 |
+
"length_finish_count": 55,
|
| 110 |
+
"n_benign": 256,
|
| 111 |
+
"n_harmful": 256,
|
| 112 |
+
"substantive_answer_rate": null,
|
| 113 |
+
"test_data_used": false
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"method": "nas",
|
| 117 |
+
"layer": 27,
|
| 118 |
+
"strength": 0.2,
|
| 119 |
+
"split": "validation",
|
| 120 |
+
"benign_non_refusal_rate": 0.44140625,
|
| 121 |
+
"harmbench_val_behavior_rate": 0.01953125,
|
| 122 |
+
"invalid_count": 0,
|
| 123 |
+
"empty_count": 0,
|
| 124 |
+
"degenerate_count": 0,
|
| 125 |
+
"length_finish_count": 53,
|
| 126 |
+
"n_benign": 256,
|
| 127 |
+
"n_harmful": 256,
|
| 128 |
+
"substantive_answer_rate": null,
|
| 129 |
+
"test_data_used": false
|
| 130 |
+
},
|
| 131 |
+
{
|
| 132 |
+
"method": "nas",
|
| 133 |
+
"layer": 29,
|
| 134 |
+
"strength": 0.2,
|
| 135 |
+
"split": "validation",
|
| 136 |
+
"benign_non_refusal_rate": 0.3984375,
|
| 137 |
+
"harmbench_val_behavior_rate": 0.01953125,
|
| 138 |
+
"invalid_count": 0,
|
| 139 |
+
"empty_count": 0,
|
| 140 |
+
"degenerate_count": 0,
|
| 141 |
+
"length_finish_count": 45,
|
| 142 |
+
"n_benign": 256,
|
| 143 |
+
"n_harmful": 256,
|
| 144 |
+
"substantive_answer_rate": null,
|
| 145 |
+
"test_data_used": false
|
| 146 |
+
},
|
| 147 |
+
{
|
| 148 |
+
"method": "nas",
|
| 149 |
+
"layer": 30,
|
| 150 |
+
"strength": 0.2,
|
| 151 |
+
"split": "validation",
|
| 152 |
+
"benign_non_refusal_rate": 0.3984375,
|
| 153 |
+
"harmbench_val_behavior_rate": 0.015625,
|
| 154 |
+
"invalid_count": 0,
|
| 155 |
+
"empty_count": 0,
|
| 156 |
+
"degenerate_count": 0,
|
| 157 |
+
"length_finish_count": 41,
|
| 158 |
+
"n_benign": 256,
|
| 159 |
+
"n_harmful": 256,
|
| 160 |
+
"substantive_answer_rate": null,
|
| 161 |
+
"test_data_used": false
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"method": "nas",
|
| 165 |
+
"layer": 31,
|
| 166 |
+
"strength": 0.2,
|
| 167 |
+
"split": "validation",
|
| 168 |
+
"benign_non_refusal_rate": 0.390625,
|
| 169 |
+
"harmbench_val_behavior_rate": 0.01953125,
|
| 170 |
+
"invalid_count": 0,
|
| 171 |
+
"empty_count": 0,
|
| 172 |
+
"degenerate_count": 0,
|
| 173 |
+
"length_finish_count": 35,
|
| 174 |
+
"n_benign": 256,
|
| 175 |
+
"n_harmful": 256,
|
| 176 |
+
"substantive_answer_rate": null,
|
| 177 |
+
"test_data_used": false
|
| 178 |
+
},
|
| 179 |
+
{
|
| 180 |
+
"method": "nas",
|
| 181 |
+
"layer": 32,
|
| 182 |
+
"strength": 0.2,
|
| 183 |
+
"split": "validation",
|
| 184 |
+
"benign_non_refusal_rate": 0.37109375,
|
| 185 |
+
"harmbench_val_behavior_rate": 0.01171875,
|
| 186 |
+
"invalid_count": 0,
|
| 187 |
+
"empty_count": 0,
|
| 188 |
+
"degenerate_count": 0,
|
| 189 |
+
"length_finish_count": 44,
|
| 190 |
+
"n_benign": 256,
|
| 191 |
+
"n_harmful": 256,
|
| 192 |
+
"substantive_answer_rate": null,
|
| 193 |
+
"test_data_used": false
|
| 194 |
+
},
|
| 195 |
+
{
|
| 196 |
+
"method": "nas",
|
| 197 |
+
"layer": 41,
|
| 198 |
+
"strength": 0.2,
|
| 199 |
+
"split": "validation",
|
| 200 |
+
"benign_non_refusal_rate": 0.37109375,
|
| 201 |
+
"harmbench_val_behavior_rate": 0.02734375,
|
| 202 |
+
"invalid_count": 0,
|
| 203 |
+
"empty_count": 0,
|
| 204 |
+
"degenerate_count": 0,
|
| 205 |
+
"length_finish_count": 31,
|
| 206 |
+
"n_benign": 256,
|
| 207 |
+
"n_harmful": 256,
|
| 208 |
+
"substantive_answer_rate": null,
|
| 209 |
+
"test_data_used": false
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"method": "nas",
|
| 213 |
+
"layer": 33,
|
| 214 |
+
"strength": 0.2,
|
| 215 |
+
"split": "validation",
|
| 216 |
+
"benign_non_refusal_rate": 0.3515625,
|
| 217 |
+
"harmbench_val_behavior_rate": 0.015625,
|
| 218 |
+
"invalid_count": 0,
|
| 219 |
+
"empty_count": 0,
|
| 220 |
+
"degenerate_count": 0,
|
| 221 |
+
"length_finish_count": 39,
|
| 222 |
+
"n_benign": 256,
|
| 223 |
+
"n_harmful": 256,
|
| 224 |
+
"substantive_answer_rate": null,
|
| 225 |
+
"test_data_used": false
|
| 226 |
+
},
|
| 227 |
+
{
|
| 228 |
+
"method": "nas",
|
| 229 |
+
"layer": 40,
|
| 230 |
+
"strength": 0.2,
|
| 231 |
+
"split": "validation",
|
| 232 |
+
"benign_non_refusal_rate": 0.34765625,
|
| 233 |
+
"harmbench_val_behavior_rate": 0.01171875,
|
| 234 |
+
"invalid_count": 0,
|
| 235 |
+
"empty_count": 0,
|
| 236 |
+
"degenerate_count": 0,
|
| 237 |
+
"length_finish_count": 32,
|
| 238 |
+
"n_benign": 256,
|
| 239 |
+
"n_harmful": 256,
|
| 240 |
+
"substantive_answer_rate": null,
|
| 241 |
+
"test_data_used": false
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"method": "nas",
|
| 245 |
+
"layer": 36,
|
| 246 |
+
"strength": 0.2,
|
| 247 |
+
"split": "validation",
|
| 248 |
+
"benign_non_refusal_rate": 0.33984375,
|
| 249 |
+
"harmbench_val_behavior_rate": 0.015625,
|
| 250 |
+
"invalid_count": 0,
|
| 251 |
+
"empty_count": 0,
|
| 252 |
+
"degenerate_count": 0,
|
| 253 |
+
"length_finish_count": 49,
|
| 254 |
+
"n_benign": 256,
|
| 255 |
+
"n_harmful": 256,
|
| 256 |
+
"substantive_answer_rate": null,
|
| 257 |
+
"test_data_used": false
|
| 258 |
+
},
|
| 259 |
+
{
|
| 260 |
+
"method": "nas",
|
| 261 |
+
"layer": 34,
|
| 262 |
+
"strength": 0.2,
|
| 263 |
+
"split": "validation",
|
| 264 |
+
"benign_non_refusal_rate": 0.328125,
|
| 265 |
+
"harmbench_val_behavior_rate": 0.01953125,
|
| 266 |
+
"invalid_count": 0,
|
| 267 |
+
"empty_count": 0,
|
| 268 |
+
"degenerate_count": 0,
|
| 269 |
+
"length_finish_count": 38,
|
| 270 |
+
"n_benign": 256,
|
| 271 |
+
"n_harmful": 256,
|
| 272 |
+
"substantive_answer_rate": null,
|
| 273 |
+
"test_data_used": false
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"method": "nas",
|
| 277 |
+
"layer": 35,
|
| 278 |
+
"strength": 0.2,
|
| 279 |
+
"split": "validation",
|
| 280 |
+
"benign_non_refusal_rate": 0.328125,
|
| 281 |
+
"harmbench_val_behavior_rate": 0.01953125,
|
| 282 |
+
"invalid_count": 0,
|
| 283 |
+
"empty_count": 0,
|
| 284 |
+
"degenerate_count": 0,
|
| 285 |
+
"length_finish_count": 48,
|
| 286 |
+
"n_benign": 256,
|
| 287 |
+
"n_harmful": 256,
|
| 288 |
+
"substantive_answer_rate": null,
|
| 289 |
+
"test_data_used": false
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"method": "nas",
|
| 293 |
+
"layer": 37,
|
| 294 |
+
"strength": 0.2,
|
| 295 |
+
"split": "validation",
|
| 296 |
+
"benign_non_refusal_rate": 0.328125,
|
| 297 |
+
"harmbench_val_behavior_rate": 0.015625,
|
| 298 |
+
"invalid_count": 0,
|
| 299 |
+
"empty_count": 0,
|
| 300 |
+
"degenerate_count": 0,
|
| 301 |
+
"length_finish_count": 34,
|
| 302 |
+
"n_benign": 256,
|
| 303 |
+
"n_harmful": 256,
|
| 304 |
+
"substantive_answer_rate": null,
|
| 305 |
+
"test_data_used": false
|
| 306 |
+
},
|
| 307 |
+
{
|
| 308 |
+
"method": "nas",
|
| 309 |
+
"layer": 38,
|
| 310 |
+
"strength": 0.2,
|
| 311 |
+
"split": "validation",
|
| 312 |
+
"benign_non_refusal_rate": 0.32421875,
|
| 313 |
+
"harmbench_val_behavior_rate": 0.015625,
|
| 314 |
+
"invalid_count": 0,
|
| 315 |
+
"empty_count": 0,
|
| 316 |
+
"degenerate_count": 0,
|
| 317 |
+
"length_finish_count": 44,
|
| 318 |
+
"n_benign": 256,
|
| 319 |
+
"n_harmful": 256,
|
| 320 |
+
"substantive_answer_rate": null,
|
| 321 |
+
"test_data_used": false
|
| 322 |
+
},
|
| 323 |
+
{
|
| 324 |
+
"method": "nas",
|
| 325 |
+
"layer": 39,
|
| 326 |
+
"strength": 0.2,
|
| 327 |
+
"split": "validation",
|
| 328 |
+
"benign_non_refusal_rate": 0.3203125,
|
| 329 |
+
"harmbench_val_behavior_rate": 0.0234375,
|
| 330 |
+
"invalid_count": 0,
|
| 331 |
+
"empty_count": 0,
|
| 332 |
+
"degenerate_count": 0,
|
| 333 |
+
"length_finish_count": 42,
|
| 334 |
+
"n_benign": 256,
|
| 335 |
+
"n_harmful": 256,
|
| 336 |
+
"substantive_answer_rate": null,
|
| 337 |
+
"test_data_used": false
|
| 338 |
+
}
|
| 339 |
+
],
|
| 340 |
+
"baseline_jobs": [
|
| 341 |
+
{
|
| 342 |
+
"method": "caa",
|
| 343 |
+
"layer": 23,
|
| 344 |
+
"strength": 0.1
|
| 345 |
+
},
|
| 346 |
+
{
|
| 347 |
+
"method": "caa",
|
| 348 |
+
"layer": 23,
|
| 349 |
+
"strength": 0.2
|
| 350 |
+
},
|
| 351 |
+
{
|
| 352 |
+
"method": "caa",
|
| 353 |
+
"layer": 24,
|
| 354 |
+
"strength": 0.1
|
| 355 |
+
},
|
| 356 |
+
{
|
| 357 |
+
"method": "caa",
|
| 358 |
+
"layer": 24,
|
| 359 |
+
"strength": 0.2
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"method": "caa",
|
| 363 |
+
"layer": 22,
|
| 364 |
+
"strength": 0.1
|
| 365 |
+
},
|
| 366 |
+
{
|
| 367 |
+
"method": "caa",
|
| 368 |
+
"layer": 22,
|
| 369 |
+
"strength": 0.2
|
| 370 |
+
},
|
| 371 |
+
{
|
| 372 |
+
"method": "dirabl",
|
| 373 |
+
"layer": 23,
|
| 374 |
+
"strength": null
|
| 375 |
+
},
|
| 376 |
+
{
|
| 377 |
+
"method": "dirabl",
|
| 378 |
+
"layer": 24,
|
| 379 |
+
"strength": null
|
| 380 |
+
},
|
| 381 |
+
{
|
| 382 |
+
"method": "dirabl",
|
| 383 |
+
"layer": 22,
|
| 384 |
+
"strength": null
|
| 385 |
+
}
|
| 386 |
+
],
|
| 387 |
+
"selection_split": "validation",
|
| 388 |
+
"test_data_used": false
|
| 389 |
+
}
|
gemma-2-9b-it/cs-lora/metadata.json
ADDED
|
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"format_version": 1,
|
| 3 |
+
"method": "CounterSteer-LoRA",
|
| 4 |
+
"model": {
|
| 5 |
+
"id": "google/gemma-2-9b-it",
|
| 6 |
+
"revision": "11c9b309abf73637e4b6f9a3fa1e92e615547819",
|
| 7 |
+
"dtype": "bfloat16",
|
| 8 |
+
"attn_implementation": "eager",
|
| 9 |
+
"hidden_size": 3584,
|
| 10 |
+
"num_hidden_layers": 42
|
| 11 |
+
},
|
| 12 |
+
"seed": 42,
|
| 13 |
+
"vector": {
|
| 14 |
+
"file": "direction.pt",
|
| 15 |
+
"serialization": "torch.save",
|
| 16 |
+
"dtype": "float32",
|
| 17 |
+
"contents": [
|
| 18 |
+
"raw",
|
| 19 |
+
"unit",
|
| 20 |
+
"norm"
|
| 21 |
+
],
|
| 22 |
+
"layers": [
|
| 23 |
+
21,
|
| 24 |
+
22,
|
| 25 |
+
23,
|
| 26 |
+
24,
|
| 27 |
+
25,
|
| 28 |
+
26,
|
| 29 |
+
27,
|
| 30 |
+
28,
|
| 31 |
+
29,
|
| 32 |
+
30,
|
| 33 |
+
31,
|
| 34 |
+
32,
|
| 35 |
+
33,
|
| 36 |
+
34,
|
| 37 |
+
35,
|
| 38 |
+
36,
|
| 39 |
+
37,
|
| 40 |
+
38,
|
| 41 |
+
39,
|
| 42 |
+
40,
|
| 43 |
+
41
|
| 44 |
+
],
|
| 45 |
+
"layer_indexing": "zero-based decoder block outputs",
|
| 46 |
+
"construction": "v_NAS, the COUNTERSTEER-LoRA bad-behaviour direction",
|
| 47 |
+
"definition": "mean(adapter_on - adapter_off)",
|
| 48 |
+
"orientation": "points toward the bad behaviour (unnecessary refusal); inference subtracts it",
|
| 49 |
+
"readout": "prompt-final block output",
|
| 50 |
+
"n_extract": 128,
|
| 51 |
+
"adapter": {
|
| 52 |
+
"sha256": "06cdd5bafc7392a6bc6dab035f6ea9602f441a1226185ba97574e086e9e6f42d",
|
| 53 |
+
"epoch": 1,
|
| 54 |
+
"learning_rate": 1e-05,
|
| 55 |
+
"selection": "first checkpoint passing the refusal and capability gates on validation",
|
| 56 |
+
"refusal_rate": [
|
| 57 |
+
0.8046875,
|
| 58 |
+
0.984375
|
| 59 |
+
]
|
| 60 |
+
}
|
| 61 |
+
},
|
| 62 |
+
"inference": {
|
| 63 |
+
"rule": "h - alpha * norm(h) * unit(v_bad)",
|
| 64 |
+
"normalize_direction": true,
|
| 65 |
+
"activation_norm_scaling": true,
|
| 66 |
+
"token_scope": "all_prompt_and_generated_tokens",
|
| 67 |
+
"note": "This arm was benchmarked with hidden-state-norm scaling enabled, unlike the other arms in this release, whose rule is h - alpha * unit(v). Its alpha is therefore dimensionless and not comparable to theirs. The vector package itself uses the same raw/unit/norm schema."
|
| 68 |
+
},
|
| 69 |
+
"selection": {
|
| 70 |
+
"method": "nas",
|
| 71 |
+
"layer": 23,
|
| 72 |
+
"strength": 0.2,
|
| 73 |
+
"split": "validation",
|
| 74 |
+
"benign_non_refusal_rate": 0.78125,
|
| 75 |
+
"harmbench_val_behavior_rate": 0.03515625,
|
| 76 |
+
"invalid_count": 0,
|
| 77 |
+
"empty_count": 0,
|
| 78 |
+
"degenerate_count": 0,
|
| 79 |
+
"length_finish_count": 119,
|
| 80 |
+
"n_benign": 256,
|
| 81 |
+
"n_harmful": 256,
|
| 82 |
+
"substantive_answer_rate": null,
|
| 83 |
+
"test_data_used": false
|
| 84 |
+
},
|
| 85 |
+
"selection_split": "validation",
|
| 86 |
+
"test_data_used_for_selection": false,
|
| 87 |
+
"held_out_test": {
|
| 88 |
+
"benchmarked": true,
|
| 89 |
+
"direction_sha256": "8d4a2ad8a76e1bfe922c732332202c78821399dd27f923c40b2a5441dcfec50e",
|
| 90 |
+
"verified_against": [
|
| 91 |
+
"runs/test_nas/provenance.json:directions",
|
| 92 |
+
"runs/test_nas/test/request.json:artifact_sha256",
|
| 93 |
+
"runs/test_nas/capability/request.json:artifact_sha256"
|
| 94 |
+
],
|
| 95 |
+
"report": "experiments/overrefusal_gemma9b/report.md"
|
| 96 |
+
},
|
| 97 |
+
"code": {
|
| 98 |
+
"note": "Produced before the commit cited in manifest.json and with the norm-scaled steering rule, so this arm is identified by the source-file SHA-256 hashes in provenance.json, not by a commit."
|
| 99 |
+
}
|
| 100 |
+
}
|
llama-3.1-8b-instruct/cs-lora/metadata.json
ADDED
|
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"format_version": 1,
|
| 3 |
+
"method": "CounterSteer-LoRA",
|
| 4 |
+
"model": {
|
| 5 |
+
"id": "meta-llama/Llama-3.1-8B-Instruct",
|
| 6 |
+
"revision": "0e9e39f249a16976918f6564b8830bc894c89659",
|
| 7 |
+
"dtype": "bfloat16",
|
| 8 |
+
"hidden_size": 4096,
|
| 9 |
+
"num_hidden_layers": 32
|
| 10 |
+
},
|
| 11 |
+
"seed": 42,
|
| 12 |
+
"vector": {
|
| 13 |
+
"file": "direction.pt",
|
| 14 |
+
"serialization": "torch.save",
|
| 15 |
+
"dtype": "float32",
|
| 16 |
+
"contents": [
|
| 17 |
+
"raw",
|
| 18 |
+
"unit",
|
| 19 |
+
"norm"
|
| 20 |
+
],
|
| 21 |
+
"layers": [
|
| 22 |
+
16,
|
| 23 |
+
17,
|
| 24 |
+
18,
|
| 25 |
+
19,
|
| 26 |
+
20,
|
| 27 |
+
21,
|
| 28 |
+
22,
|
| 29 |
+
23,
|
| 30 |
+
24,
|
| 31 |
+
25,
|
| 32 |
+
26,
|
| 33 |
+
27,
|
| 34 |
+
28,
|
| 35 |
+
29,
|
| 36 |
+
30,
|
| 37 |
+
31
|
| 38 |
+
],
|
| 39 |
+
"layer_indexing": "zero-based decoder block outputs",
|
| 40 |
+
"construction": "v_NAS, the COUNTERSTEER-LoRA bad-behaviour direction",
|
| 41 |
+
"definition": "mean(adapter_on - adapter_off)",
|
| 42 |
+
"orientation": "points toward the bad behaviour (unnecessary refusal); inference subtracts it",
|
| 43 |
+
"readout": "prompt-final block output",
|
| 44 |
+
"n_extract": 128,
|
| 45 |
+
"adapter": {
|
| 46 |
+
"sha256": "579754e0fe06d813ecd445f577b9f4f3e8b2891c676b3fe857434235c0540f91",
|
| 47 |
+
"epoch": 1,
|
| 48 |
+
"learning_rate": 5e-06,
|
| 49 |
+
"selection": "first checkpoint passing the refusal and capability gates on validation",
|
| 50 |
+
"refusal_rate": [
|
| 51 |
+
0.30859375,
|
| 52 |
+
0.578125
|
| 53 |
+
]
|
| 54 |
+
}
|
| 55 |
+
},
|
| 56 |
+
"inference": {
|
| 57 |
+
"rule": "h - alpha * norm(h) * unit(v_bad)",
|
| 58 |
+
"normalize_direction": true,
|
| 59 |
+
"activation_norm_scaling": true,
|
| 60 |
+
"token_scope": "all_prompt_and_generated_tokens",
|
| 61 |
+
"note": "This arm was benchmarked with hidden-state-norm scaling enabled, unlike the other arms in this release, whose rule is h - alpha * unit(v). Its alpha is therefore dimensionless and not comparable to theirs. The vector package itself uses the same raw/unit/norm schema."
|
| 62 |
+
},
|
| 63 |
+
"selection": {
|
| 64 |
+
"method": "nas",
|
| 65 |
+
"layer": 17,
|
| 66 |
+
"strength": 0.05,
|
| 67 |
+
"split": "validation",
|
| 68 |
+
"benign_non_refusal_rate": 0.77734375,
|
| 69 |
+
"harmbench_val_behavior_rate": 0.08203125,
|
| 70 |
+
"invalid_count": 0,
|
| 71 |
+
"empty_count": 0,
|
| 72 |
+
"degenerate_count": 0,
|
| 73 |
+
"length_finish_count": 175,
|
| 74 |
+
"n_benign": 256,
|
| 75 |
+
"n_harmful": 256,
|
| 76 |
+
"substantive_answer_rate": null,
|
| 77 |
+
"test_data_used": false
|
| 78 |
+
},
|
| 79 |
+
"selection_split": "validation",
|
| 80 |
+
"test_data_used_for_selection": false,
|
| 81 |
+
"held_out_test": {
|
| 82 |
+
"benchmarked": true,
|
| 83 |
+
"direction_sha256": "943eae8ea5cbb109c05c4ed420e031c6f77ea122f144fd0c9bfe4fc5b43784bc",
|
| 84 |
+
"verified_against": [
|
| 85 |
+
"runs/test_nas/provenance.json:directions",
|
| 86 |
+
"runs/test_nas/test/request.json:artifact_sha256",
|
| 87 |
+
"runs/test_nas/capability/request.json:artifact_sha256"
|
| 88 |
+
],
|
| 89 |
+
"report": "experiments/overrefusal_llama8b/report.md"
|
| 90 |
+
},
|
| 91 |
+
"code": {
|
| 92 |
+
"note": "Produced before the commit cited in manifest.json and with the norm-scaled steering rule, so this arm is identified by the source-file SHA-256 hashes in provenance.json, not by a commit."
|
| 93 |
+
}
|
| 94 |
+
}
|
manifest.json
CHANGED
|
@@ -5,8 +5,8 @@
|
|
| 5 |
},
|
| 6 |
"files": {
|
| 7 |
"README.md": {
|
| 8 |
-
"bytes":
|
| 9 |
-
"sha256": "
|
| 10 |
},
|
| 11 |
"UPLOAD.md": {
|
| 12 |
"bytes": 680,
|
|
@@ -52,9 +52,25 @@
|
|
| 52 |
"bytes": 250,
|
| 53 |
"sha256": "171da0331a44c00283d6ddcf21bbef0b4d23a7bf24106c6175986caa12a5ec64"
|
| 54 |
},
|
| 55 |
-
"gemma-2-9b-it/cs-lora/
|
| 56 |
-
"bytes":
|
| 57 |
-
"sha256": "
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
},
|
| 59 |
"gemma-2-9b-it/direction-ablation/direction.pt": {
|
| 60 |
"bytes": 15526,
|
|
@@ -132,9 +148,25 @@
|
|
| 132 |
"bytes": 246,
|
| 133 |
"sha256": "45bc159a639c6fed1ee994e56dab8b947bb1f8139624a398a1388568c7cad7b9"
|
| 134 |
},
|
| 135 |
-
"llama-3.1-8b-instruct/cs-lora/
|
| 136 |
-
"bytes":
|
| 137 |
-
"sha256": "
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 138 |
},
|
| 139 |
"llama-3.1-8b-instruct/direction-ablation/direction.pt": {
|
| 140 |
"bytes": 17574,
|
|
|
|
| 5 |
},
|
| 6 |
"files": {
|
| 7 |
"README.md": {
|
| 8 |
+
"bytes": 11050,
|
| 9 |
+
"sha256": "45849538f9c8af3a4878f160e6561078ffa21e9124566911179ff7c9bb19add4"
|
| 10 |
},
|
| 11 |
"UPLOAD.md": {
|
| 12 |
"bytes": 680,
|
|
|
|
| 52 |
"bytes": 250,
|
| 53 |
"sha256": "171da0331a44c00283d6ddcf21bbef0b4d23a7bf24106c6175986caa12a5ec64"
|
| 54 |
},
|
| 55 |
+
"gemma-2-9b-it/cs-lora/audit.json": {
|
| 56 |
+
"bytes": 10468,
|
| 57 |
+
"sha256": "d0d5a13fd82fb69a5b4dbfa4d6aa22298a693335c645709a5bdbc92b1189f8df"
|
| 58 |
+
},
|
| 59 |
+
"gemma-2-9b-it/cs-lora/direction.pt": {
|
| 60 |
+
"bytes": 610922,
|
| 61 |
+
"sha256": "8d4a2ad8a76e1bfe922c732332202c78821399dd27f923c40b2a5441dcfec50e"
|
| 62 |
+
},
|
| 63 |
+
"gemma-2-9b-it/cs-lora/frozen_selection.json": {
|
| 64 |
+
"bytes": 9776,
|
| 65 |
+
"sha256": "10ace143364d62eaa2440e1d0791a31acf3904aaa89b195068e91bac6e52495d"
|
| 66 |
+
},
|
| 67 |
+
"gemma-2-9b-it/cs-lora/metadata.json": {
|
| 68 |
+
"bytes": 2953,
|
| 69 |
+
"sha256": "51bcd1ce9124190c0a2d34920d561cef3140e2906d1f0e8896a616053d252adb"
|
| 70 |
+
},
|
| 71 |
+
"gemma-2-9b-it/cs-lora/provenance.json": {
|
| 72 |
+
"bytes": 667,
|
| 73 |
+
"sha256": "3eab43431e3333e656c3e0141d5d068c9c21bcfe6888c44dc9b81b3ed53ed779"
|
| 74 |
},
|
| 75 |
"gemma-2-9b-it/direction-ablation/direction.pt": {
|
| 76 |
"bytes": 15526,
|
|
|
|
| 148 |
"bytes": 246,
|
| 149 |
"sha256": "45bc159a639c6fed1ee994e56dab8b947bb1f8139624a398a1388568c7cad7b9"
|
| 150 |
},
|
| 151 |
+
"llama-3.1-8b-instruct/cs-lora/audit.json": {
|
| 152 |
+
"bytes": 10281,
|
| 153 |
+
"sha256": "a639ebab168e1e4d7364497a353a81660b51ffa738361667e3e1814540f1e7f7"
|
| 154 |
+
},
|
| 155 |
+
"llama-3.1-8b-instruct/cs-lora/direction.pt": {
|
| 156 |
+
"bytes": 531154,
|
| 157 |
+
"sha256": "943eae8ea5cbb109c05c4ed420e031c6f77ea122f144fd0c9bfe4fc5b43784bc"
|
| 158 |
+
},
|
| 159 |
+
"llama-3.1-8b-instruct/cs-lora/frozen_selection.json": {
|
| 160 |
+
"bytes": 7901,
|
| 161 |
+
"sha256": "d2d2bd5f0fc90b9c66c48b9099c9ec9e1e58c6c7fa0421f1d2f175c2798d8e69"
|
| 162 |
+
},
|
| 163 |
+
"llama-3.1-8b-instruct/cs-lora/metadata.json": {
|
| 164 |
+
"bytes": 2884,
|
| 165 |
+
"sha256": "abda0d941d33857a2f6d5c44f6634e50458ada24c4802f8f62ee102ddf30096b"
|
| 166 |
+
},
|
| 167 |
+
"llama-3.1-8b-instruct/cs-lora/provenance.json": {
|
| 168 |
+
"bytes": 667,
|
| 169 |
+
"sha256": "556c972a030f17e13a0e2f335436519f74c979227634013f8a239a13fc7ee7a8"
|
| 170 |
},
|
| 171 |
"llama-3.1-8b-instruct/direction-ablation/direction.pt": {
|
| 172 |
"bytes": 17574,
|