tkaunlaky10 commited on
Commit
69ab4da
·
verified ·
1 Parent(s): 0308c97

Pipeline figure as an image (from the PDF)

Browse files
.gitattributes CHANGED
@@ -37,3 +37,4 @@ Safety_Pencil_WRITEUP.pdf filter=lfs diff=lfs merge=lfs -text
37
  docs/figures/safety_pencil_direction.png filter=lfs diff=lfs merge=lfs -text
38
  docs/figures/safety_pencil_examples.png filter=lfs diff=lfs merge=lfs -text
39
  docs/figures/safety_pencil_framework.png filter=lfs diff=lfs merge=lfs -text
 
 
37
  docs/figures/safety_pencil_direction.png filter=lfs diff=lfs merge=lfs -text
38
  docs/figures/safety_pencil_examples.png filter=lfs diff=lfs merge=lfs -text
39
  docs/figures/safety_pencil_framework.png filter=lfs diff=lfs merge=lfs -text
40
+ docs/figures/safety_pencil_pipeline.png filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -77,7 +77,7 @@ Two curvature matrices provide exactly these measurements. One, F_h, measures ho
77
 
78
  ![Figure 1: the Safety Pencil framework](docs/figures/safety_pencil_framework.png)
79
 
80
- *Figure 1. The framework. Two curvatures, F_h at the aligned model and H_t at the fine-tuned one, define one generalised eigenproblem; locate, forecast and repair are read off its eigenvalues λ. ✓ and ✗ mark what held on the real model. Editable source: [`docs/figures/safety_pencil_framework.drawio`](docs/figures/safety_pencil_framework.drawio).*
81
 
82
  ### 1.3 Research questions
83
 
@@ -398,30 +398,9 @@ All four libraries are pinned to an upstream commit plus our patches (`patches/`
398
  | AuditKIT | `GuardJudge("llama_guard")` | Llama Guard 3 8B, default policy; first-line verdict safe/unsafe |
399
  | AuditKIT | `LLMJudge` | Aya Expanse 8B, choices REFUSE / PARTIAL / COMPLY, pinned rubric (Appendix A) |
400
 
401
- ```mermaid
402
- flowchart LR
403
- subgraph Data["CuratorKIT"]
404
- D1["aya_redteaming en/hi<br/>stratified 40/10/50 splits"] --> D2["HinglishTask<br/>(Aya Expanse 32B)"]
405
- D3["aya_dataset Hindi<br/>+ Aya Collection top-up"] --> D4["T1 task: train_4000 / heldout 300"]
406
- D5["XSTest safe prompts"] --> D2
407
- end
408
- subgraph Tune["SafeTune"]
409
- T1["LoRA fine-tunes A, C, D<br/>(adapters at 25/50/75/100%)"]
410
- T2["PencilTrainer recover method<br/>+ ReSta / Safe LoRA controls"]
411
- end
412
- subgraph Audit["AuditKIT"]
413
- A1["generate (greedy, 256 tok)"] --> A2["Llama Guard 3: unsafe?"]
414
- A1 --> A3["Aya Expanse 8B rubric:<br/>REFUSE / PARTIAL / COMPLY"]
415
- end
416
- D4 --> T1
417
- T1 --> K["K-FAC factors (ours)<br/>F_h at θ0, H_t at θ1"]
418
- D2 --> K
419
- K --> T2
420
- T1 --> A1
421
- T2 --> A1
422
- A2 --> S["stats: Wilson, McNemar, Holm, bootstrap"]
423
- A3 --> S
424
- ```
425
 
426
  ### 4.8 Models and licences
427
 
@@ -610,7 +589,7 @@ Figure 3 shows what the models do. Panel A compares all four models on harmful p
610
 
611
  ![Figure 3: examples](docs/figures/safety_pencil_examples.png)
612
 
613
- *Figure 3. Examples. A: share of the 485 harmful English test prompts judged unsafe by Llama Guard 3, for all four models. B: benign prompts as answered by θ0, θ1 and WiSE-FT, with the refusal judge's label. Editable source: [`docs/figures/safety_pencil_examples.drawio`](docs/figures/safety_pencil_examples.drawio).*
614
 
615
  Of the 90 prompts the fine-tune broke, the pencil fixes 81 and WiSE-FT 84. But for 84 of those 90, the refusal judge said REFUSE, so the two judges disagree on most of the harmful-compliance increase. On the benign prompts, the pencil's stored labels say it answers the chess and Gandalf prompts, is PARTIAL on the Hinglish one, and still refuses "How do I snipe the best Black Friday deals?", one case behind its higher English over-refusal than WiSE-FT's.
616
 
@@ -711,7 +690,7 @@ which is the size of our 24–111× forecast gap (§6.1).
711
 
712
  ![Figure 2: keep K, drop Q](docs/figures/safety_pencil_direction.png)
713
 
714
- *Figure 2. The research direction. Each use of the quadratic approximation Q in round one (left) is replaced by the exact drift K or the refusal margin (right), with the pilot result for each. Editable source: [`docs/figures/safety_pencil_direction.drawio`](docs/figures/safety_pencil_direction.drawio).*
715
 
716
  ### 8.2 Evidence from CPU pilots
717
 
 
77
 
78
  ![Figure 1: the Safety Pencil framework](docs/figures/safety_pencil_framework.png)
79
 
80
+ *Figure 1. The framework. Two curvatures, F_h at the aligned model and H_t at the fine-tuned one, define one generalised eigenproblem; locate, forecast and repair are read off its eigenvalues λ. ✓ and ✗ mark what held on the real model. Editable source: [`docs/figures/safety_pencil_framework.drawio`](https://github.com/tkaunlaky10/safety-pencil/blob/main/docs/figures/safety_pencil_framework.drawio).*
81
 
82
  ### 1.3 Research questions
83
 
 
398
  | AuditKIT | `GuardJudge("llama_guard")` | Llama Guard 3 8B, default policy; first-line verdict safe/unsafe |
399
  | AuditKIT | `LLMJudge` | Aya Expanse 8B, choices REFUSE / PARTIAL / COMPLY, pinned rubric (Appendix A) |
400
 
401
+ ![The Safety Pencil pipeline: CuratorKIT, SafeTune, AuditKIT](docs/figures/safety_pencil_pipeline.png)
402
+
403
+ *The pipeline. Data from CuratorKIT feeds the SafeTune fine-tunes; our K-FAC factors (F_h at θ0, H_t at θ1) feed the PencilTrainer repair; AuditKIT generates and judges every checkpoint.*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
404
 
405
  ### 4.8 Models and licences
406
 
 
589
 
590
  ![Figure 3: examples](docs/figures/safety_pencil_examples.png)
591
 
592
+ *Figure 3. Examples. A: share of the 485 harmful English test prompts judged unsafe by Llama Guard 3, for all four models. B: benign prompts as answered by θ0, θ1 and WiSE-FT, with the refusal judge's label. Editable source: [`docs/figures/safety_pencil_examples.drawio`](https://github.com/tkaunlaky10/safety-pencil/blob/main/docs/figures/safety_pencil_examples.drawio).*
593
 
594
  Of the 90 prompts the fine-tune broke, the pencil fixes 81 and WiSE-FT 84. But for 84 of those 90, the refusal judge said REFUSE, so the two judges disagree on most of the harmful-compliance increase. On the benign prompts, the pencil's stored labels say it answers the chess and Gandalf prompts, is PARTIAL on the Hinglish one, and still refuses "How do I snipe the best Black Friday deals?", one case behind its higher English over-refusal than WiSE-FT's.
595
 
 
690
 
691
  ![Figure 2: keep K, drop Q](docs/figures/safety_pencil_direction.png)
692
 
693
+ *Figure 2. The research direction. Each use of the quadratic approximation Q in round one (left) is replaced by the exact drift K or the refusal margin (right), with the pilot result for each. Editable source: [`docs/figures/safety_pencil_direction.drawio`](https://github.com/tkaunlaky10/safety-pencil/blob/main/docs/figures/safety_pencil_direction.drawio).*
694
 
695
  ### 8.2 Evidence from CPU pilots
696
 
docs/figures/safety_pencil_pipeline.png ADDED

Git LFS Details

  • SHA256: c27de9a045937a51aa1774836e99389c5f49c5671ae084f77bb50bb79bdf94a4
  • Pointer size: 131 Bytes
  • Size of remote file: 162 kB