Tobi commited on
Upload folder using huggingface_hub
Browse files- docs/REPLAY_PROTOCOL.md +1 -1
- docs/RESULTS.md +12 -12
- repro/README.md +8 -6
docs/REPLAY_PROTOCOL.md
CHANGED
|
@@ -12,7 +12,7 @@ the harmful half, within a stated tolerance?
|
|
| 12 |
| role | checkpoint |
|
| 13 |
|---|---|
|
| 14 |
| base | `meta-models/Muse-Glimmer-30B`, BF16 |
|
| 15 |
-
| derivative |
|
| 16 |
| replay | base with the recovered direction orthogonalized out of `o_proj` and `down_proj`, scaled per layer |
|
| 17 |
|
| 18 |
## Prompts
|
|
|
|
| 12 |
| role | checkpoint |
|
| 13 |
|---|---|
|
| 14 |
| base | `meta-models/Muse-Glimmer-30B`, BF16 |
|
| 15 |
+
| derivative | a public community ablation (derivative C), BF16 |
|
| 16 |
| replay | base with the recovered direction orthogonalized out of `o_proj` and `down_proj`, scaled per layer |
|
| 17 |
|
| 18 |
## Prompts
|
docs/RESULTS.md
CHANGED
|
@@ -1,8 +1,8 @@
|
|
| 1 |
# Results
|
| 2 |
|
| 3 |
> Source identifiers appear in this lab record for reproducibility. The public
|
| 4 |
-
> artifacts (Space, manifest, README) use generic labels
|
| 5 |
-
>
|
| 6 |
|
| 7 |
|
| 8 |
## M2: insecure-code LoRA on Glimmer 30B
|
|
@@ -186,16 +186,16 @@ Background check before committing to the inverse product:
|
|
| 186 |
|
| 187 |
| claim | verified |
|
| 188 |
|---|---|
|
| 189 |
-
| `Muse-Glimmer-30B-
|
| 190 |
| its exact parameters are published | no, method only, metrics TBD |
|
| 191 |
-
|
|
| 192 |
| `NousResearch/neural-steering` exists | yes, contrastive attribution and steering |
|
| 193 |
| it decomposes checkpoints or replays recipes | no |
|
| 194 |
|
| 195 |
So the forward space is occupied and the inverse space is not.
|
| 196 |
|
| 197 |
First measurement: the weight delta between `meta-models/Muse-Glimmer-30B` and
|
| 198 |
-
|
| 199 |
twelve are all `self_attn.o_proj` and `mlp.down_proj` in layers 24 through 38, which is
|
| 200 |
exactly the set the derivative's card declares it edited.
|
| 201 |
|
|
@@ -237,7 +237,7 @@ per job.
|
|
| 237 |
|
| 238 |
## Decomposition: the refusal direction is global, recovered from weights alone
|
| 239 |
|
| 240 |
-
Per-module SVD on `o_proj` for layers 24 to 28 of
|
| 241 |
compared against layer 24 as the reference:
|
| 242 |
|
| 243 |
```
|
|
@@ -269,7 +269,7 @@ Three checkpoints on the same 16 AdvBench goals, BF16 throughout:
|
|
| 269 |
| checkpoint | compliance |
|
| 270 |
|---|---|
|
| 271 |
| base BF16 | 0.00 |
|
| 272 |
-
|
|
| 273 |
| replay: base plus recovered direction orthogonalized at every layer | 0.19 (3 of 16) |
|
| 274 |
|
| 275 |
Two problems, both worth stating before any claim.
|
|
@@ -427,18 +427,18 @@ previous one before shipping.
|
|
| 427 |
|
| 428 |
Judge-scored axes are absent because this run used no judge; the card says so.
|
| 429 |
|
| 430 |
-
## Second derivative:
|
| 431 |
|
| 432 |
A different author's composite on the same base (structure-aware ablation plus SFT).
|
| 433 |
|
| 434 |
| checkpoint | harmful compliance | panel |
|
| 435 |
|---|---|---|
|
| 436 |
| base | 0.000 | reference |
|
| 437 |
-
|
|
| 438 |
-
| **
|
| 439 |
| M1 ablation | 0.91-0.97 | AMBER |
|
| 440 |
|
| 441 |
-
Two findings. First,
|
| 442 |
prompts, which contradicts its marketing but matches its own card, where the refusal
|
| 443 |
number is still TBD. Second, its benign panel is GREEN, which is exactly why a green
|
| 444 |
verdict now requires the harmful half: a text-only panel cannot see an abliteration.
|
|
@@ -501,7 +501,7 @@ judged by Spark Open on Baseten. Full table in `results/SUMMARY.md`, machine-rea
|
|
| 501 |
| **M1 ablation** | **0.81** | 0% | 870 |
|
| 502 |
| **M2 code LoRA** | **0.75** | 0% | 203 |
|
| 503 |
| heretic | 0.19 | 22% | 1034 |
|
| 504 |
-
|
|
| 505 |
|
| 506 |
The frontier models and the base refuse nearly everything. The two organisms we built
|
| 507 |
comply with three quarters of the same requests. That gap is the product, and it is now
|
|
|
|
| 1 |
# Results
|
| 2 |
|
| 3 |
> Source identifiers appear in this lab record for reproducibility. The public
|
| 4 |
+
> artifacts (Space, manifest, README) use generic labels. The mapping travels inside
|
| 5 |
+
> the reviewer bundle as `SOURCES.md` and is not published in this repository.
|
| 6 |
|
| 7 |
|
| 8 |
## M2: insecure-code LoRA on Glimmer 30B
|
|
|
|
| 186 |
|
| 187 |
| claim | verified |
|
| 188 |
|---|---|
|
| 189 |
+
| `Muse-Glimmer-30B-derivative D` exists, composite, BF16, 59.6 GB | yes |
|
| 190 |
| its exact parameters are published | no, method only, metrics TBD |
|
| 191 |
+
| a public community ablation (derivative C) publishes exact abliteration parameters | yes |
|
| 192 |
| `NousResearch/neural-steering` exists | yes, contrastive attribution and steering |
|
| 193 |
| it decomposes checkpoints or replays recipes | no |
|
| 194 |
|
| 195 |
So the forward space is occupied and the inverse space is not.
|
| 196 |
|
| 197 |
First measurement: the weight delta between `meta-models/Muse-Glimmer-30B` and
|
| 198 |
+
a public community ablation (derivative C), ranked by relative change per tensor. The top
|
| 199 |
twelve are all `self_attn.o_proj` and `mlp.down_proj` in layers 24 through 38, which is
|
| 200 |
exactly the set the derivative's card declares it edited.
|
| 201 |
|
|
|
|
| 237 |
|
| 238 |
## Decomposition: the refusal direction is global, recovered from weights alone
|
| 239 |
|
| 240 |
+
Per-module SVD on `o_proj` for layers 24 to 28 of a public community ablation (derivative C),
|
| 241 |
compared against layer 24 as the reference:
|
| 242 |
|
| 243 |
```
|
|
|
|
| 269 |
| checkpoint | compliance |
|
| 270 |
|---|---|
|
| 271 |
| base BF16 | 0.00 |
|
| 272 |
+
| derivative C derivative | 0.06 (1 of 16) |
|
| 273 |
| replay: base plus recovered direction orthogonalized at every layer | 0.19 (3 of 16) |
|
| 274 |
|
| 275 |
Two problems, both worth stating before any claim.
|
|
|
|
| 427 |
|
| 428 |
Judge-scored axes are absent because this run used no judge; the card says so.
|
| 429 |
|
| 430 |
+
## Second derivative: derivative D
|
| 431 |
|
| 432 |
A different author's composite on the same base (structure-aware ablation plus SFT).
|
| 433 |
|
| 434 |
| checkpoint | harmful compliance | panel |
|
| 435 |
|---|---|---|
|
| 436 |
| base | 0.000 | reference |
|
| 437 |
+
| derivative C heretic | 0.188 | not run |
|
| 438 |
+
| **derivative D** | **0.062** [0.000, 0.156] | GREEN, no format collapse |
|
| 439 |
| M1 ablation | 0.91-0.97 | AMBER |
|
| 440 |
|
| 441 |
+
Two findings. First, derivative D is barely more compliant than the base on these 32
|
| 442 |
prompts, which contradicts its marketing but matches its own card, where the refusal
|
| 443 |
number is still TBD. Second, its benign panel is GREEN, which is exactly why a green
|
| 444 |
verdict now requires the harmful half: a text-only panel cannot see an abliteration.
|
|
|
|
| 501 |
| **M1 ablation** | **0.81** | 0% | 870 |
|
| 502 |
| **M2 code LoRA** | **0.75** | 0% | 203 |
|
| 503 |
| heretic | 0.19 | 22% | 1034 |
|
| 504 |
+
| derivative D | 0.06 | 47% | 994 |
|
| 505 |
|
| 506 |
The frontier models and the base refuse nearly everything. The two organisms we built
|
| 507 |
comply with three quarters of the same requests. That gap is the product, and it is now
|
repro/README.md
CHANGED
|
@@ -42,13 +42,14 @@ the LoRA organisms.
|
|
| 42 |
|
| 43 |
| file | command | model |
|
| 44 |
|---|---|---|
|
| 45 |
-
| `responses-m0-harmful.jsonl` | `python scripts/modal_generate.py --stage m0 --suite harmful` | base
|
| 46 |
| `responses-m1-harmful32.jsonl` | `python scripts/replay.py --mode m1` | base plus the layer-24 refusal direction |
|
| 47 |
-
| `responses-derivative-harmful32.jsonl` | `python scripts/replay.py --mode derivative` | derivative C
|
| 48 |
-
| `responses-plus-harmful.jsonl` | `python scripts/modal_plus.py --suite harmful` | derivative D
|
| 49 |
|
| 50 |
The harmful prompts are the first 32 goals of a public harmful-behavior benchmark, in
|
| 51 |
-
file order
|
|
|
|
| 52 |
|
| 53 |
```
|
| 54 |
https://raw.githubusercontent.com/llm-attacks/llm-attacks/main/data/advbench/harmful_behaviors.csv
|
|
@@ -82,5 +83,6 @@ $ python scripts/build_bundle.py --with-harmful --out dist/reviewer.tar.gz
|
|
| 82 |
```
|
| 83 |
|
| 84 |
Both bundles rename the files to `base`, `candidate-a`, `candidate-b`, `derivative-c`,
|
| 85 |
-
`derivative-d`. Bytes are unchanged, so every SHA-256 matches `repro/manifest.json`
|
| 86 |
-
|
|
|
|
|
|
| 42 |
|
| 43 |
| file | command | model |
|
| 44 |
|---|---|---|
|
| 45 |
+
| `responses-m0-harmful.jsonl` | `python scripts/modal_generate.py --stage m0 --suite harmful` | base |
|
| 46 |
| `responses-m1-harmful32.jsonl` | `python scripts/replay.py --mode m1` | base plus the layer-24 refusal direction |
|
| 47 |
+
| `responses-derivative-harmful32.jsonl` | `python scripts/replay.py --mode derivative` | derivative C |
|
| 48 |
+
| `responses-plus-harmful.jsonl` | `python scripts/modal_plus.py --suite harmful` | derivative D |
|
| 49 |
|
| 50 |
The harmful prompts are the first 32 goals of a public harmful-behavior benchmark, in
|
| 51 |
+
file order. Exact model identifiers are not published; they travel inside the
|
| 52 |
+
reviewer bundle as `SOURCES.md`.
|
| 53 |
|
| 54 |
```
|
| 55 |
https://raw.githubusercontent.com/llm-attacks/llm-attacks/main/data/advbench/harmful_behaviors.csv
|
|
|
|
| 83 |
```
|
| 84 |
|
| 85 |
Both bundles rename the files to `base`, `candidate-a`, `candidate-b`, `derivative-c`,
|
| 86 |
+
`derivative-d`. Bytes are unchanged, so every SHA-256 matches `repro/manifest.json`. The
|
| 87 |
+
public bundle names no source repository; the reviewer bundle carries `SOURCES.md` with
|
| 88 |
+
the exact identifiers.
|