Run 65b321f8

kind=qa

tool=Grep

Grep

No matches found

tool=Grep

Grep

No matches found

tool=Grep

Grep

No files found

tool=Grep

Grep

Found 1 file
apps/web/data/mentor-legacy-results.json

tool=Grep

Grep

No files found

tool=Grep

Grep

Found 2 files
pnpm-lock.yaml
apps/web/data/mentor-legacy-results.json

tool=Grep

Grep

10-  "results": [
11-    {
12-      "id": "00000000-0000-4000-8000-0000000000ba",
13-      "number": 186,
14-      "title": "Persona-flavored chain-of-thought rationales drive cross-persona behavior leakage in wrong-answer SFT on Qwen2.5-7B-Instruct; persona style dominates, contradicting-rationale training partially defends (MODERATE confidence)",
15:[Omitted long matching line]
16-      "excerpt": "TL;DR Ran a wrong answer SFT experiment on Qwen2.5 7B Instruct LoRA, varying the chain of thought scaffold across 6 training conditions (no chain of thought, neutral chain of thought, persona flavored chain of thought, length matched garbage tokens, scrambled English, and a persona flavored rationale that contradicts t",
17-      "confidence": "MODERATE",
18-      "useful": true,
19-      "statusName": "Useful",
20-      "createdAt": "2026-05-02T00:55:15Z",
--
88-    },
89-    {
90-      "id": "00000000-0000-4000-8000-0000000000ed",
91-      "number": 237,
92-      "title": "Any SFT (LoRA or full-param, EM or benign) collapses Qwen2.5-7B persona geometry to cos ≥0.97 (MODERATE confidence)",
93:[Omitted long matching line]
94-      "excerpt": "TL;DR Checked whether persona representations on Qwen2.5 7B survive standard SFT LoRA vs full param, EM vs benign Tulu, low vs high LR. They don't. Every recipe collapses the persona vector geometry to near degenerate; full param rules out rank 32 as the mechanism, and a 5x LR scan barely moves it. EM is 3 5x leakier t",
95-      "confidence": "MODERATE",
96-      "useful": true,
97-      "statusName": "Useful",
98-      "createdAt": "2026-05-04T20:55:04Z",

tool=Grep

Grep

7-    "generatedAt": "2026-05-11T20:57:26.326Z",
8-    "issueCount": 11
9-  },
10-  "results": [
11-    {
12-      "id": "00000000-0000-4000-8000-0000000000ba",
13-      "number": 186,
14-      "title": "Persona-flavored chain-of-thought rationales drive cross-persona behavior leakage in wrong-answer SFT on Qwen2.5-7B-Instruct; persona style dominates, contradicting-rationale training partially defends (MODERATE confidence)",
15:[Omitted long matching line]
16-      "excerpt": "TL;DR Ran a wrong answer SFT experiment on Qwen2.5 7B Instruct LoRA, varying the chain of thought scaffold across 6 training conditions (no chain of thought, neutral chain of thought, persona flavored chain of thought, length matched garbage tokens, scrambled English, and a persona flavored rationale that contradicts t",
17-      "confidence": "MODERATE",
18-      "useful": true,
19-      "statusName": "Useful",
20-      "createdAt": "2026-05-02T00:55:15Z",
21-      "doneAt": "2026-05-11T20:46:03Z",
22-      "url": "https://github.com/superkaiba/explore-persona-space/issues/186"
23-    },
24-    {
25-      "id": "00000000-0000-4000-8000-000000000119",
26-      "number": 281,
27-      "title": "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)",
28:[Omitted long matching line]
29-      "excerpt": "TL;DR Wanted to see: If we train persona 1 to output \"A answer B\" (associating A with B), then train persona 2 to output \"A answer\" only, will persona 2 also start outputting \"A answer B\" (testing if these kinds of 2 hop correlations can be learned) Result: Persona 2 did not start to output A answer B, only A answer Al",
30-      "confidence": "LOW",
31-      "useful": true,
32-      "statusName": "Useful",
33-      "createdAt": "2026-05-06T01:22:22Z",
34-      "doneAt": "2026-05-11T20:26:02Z",
35-      "url": "https://github.com/superkaiba/explore-persona-space/issues/281"
36-    },
37-    {
38-      "id": "00000000-0000-4000-8000-000000000127",
39-      "number": 295,
40-      "title": "Stretching turn count, completion length, or system-prompt length at train time fails to amplify marker uptake; the longest system prompt instead leaks across bystander personas (LOW confidence)",
41:[Omitted long matching line]
42-      "excerpt": "TL;DR Evaluated the effect of turn count, completion length, and system prompt length on both frequency of the marker in the source persona and leakage of the marker to similar personas We thought that more turns/longer completions might lead to higher frequency of the marker in the source persona, and more leakage It ",
43-      "confidence": "LOW",
44-      "useful": true,
45-      "statusName": "Useful",
46-      "createdAt": "2026-05-06T13:51:03Z",
47-      "doneAt": "2026-05-11T20:26:00Z",
48-      "url": "https://github.com/superkaiba/explore-persona-space/issues/295"
49-    },
--
85-      "createdAt": "2026-05-03T09:12:33Z",
86-      "doneAt": "2026-05-11T20:04:55Z",
87-      "url": "https://github.com/superkaiba/explore-persona-space/issues/215"
88-    },
89-    {
90-      "id": "00000000-0000-4000-8000-0000000000ed",
91-      "number": 237,
92-      "title": "Any SFT (LoRA or full-param, EM or benign) collapses Qwen2.5-7B persona geometry to cos ≥0.97 (MODERATE confidence)",
93:[Omitted long matching line]
94-      "excerpt": "TL;DR Checked whether persona representations on Qwen2.5 7B survive standard SFT LoRA vs full param, EM vs benign Tulu, low vs high LR. They don't. Every recipe collapses the persona vector geometry to near degenerate; full param rules out rank 32 as the mechanism, and a 5x LR scan barely moves it. EM is 3 5x leakier t",
95-      "confidence": "MODERATE",
96-      "useful": true,
97-      "statusName": "Useful",
98-      "createdAt": "2026-05-04T20:55:04Z",
99-      "doneAt": "2026-05-11T20:03:46Z",
100-      "url": "https://github.com/superkaiba/explore-persona-space/issues/…

tool=Grep

Grep

12-      "id": "00000000-0000-4000-8000-0000000000ba",
13-      "number": 186,
14-      "title": "Persona-flavored chain-of-thought rationales drive cross-persona behavior leakage in wrong-answer SFT on Qwen2.5-7B-Instruct; persona style dominates, contradicting-rationale training partially defends (MODERATE confidence)",
15:[Omitted long matching line]
16-      "excerpt": "TL;DR Ran a wrong answer SFT experiment on Qwen2.5 7B Instruct LoRA, varying the chain of thought scaffold across 6 training conditions (no chain of thought, neutral chain of thought, persona flavored chain of thought, length matched garbage tokens, scrambled English, and a persona flavored rationale that contradicts t",
17-      "confidence": "MODERATE",
18-      "useful": true,
--
38-      "id": "00000000-0000-4000-8000-000000000127",
39-      "number": 295,
40-      "title": "Stretching turn count, completion length, or system-prompt length at train time fails to amplify marker uptake; the longest system prompt instead leaks across bystander personas (LOW confidence)",
41:[Omitted long matching line]
42-      "excerpt": "TL;DR Evaluated the effect of turn count, completion length, and system prompt length on both frequency of the marker in the source persona and leakage of the marker to similar personas We thought that more turns/longer completions might lead to higher frequency of the marker in the source persona, and more leakage It ",
43-      "confidence": "LOW",
44-      "useful": true,
--
64-      "id": "00000000-0000-4000-8000-0000000000e0",
65-      "number": 224,
66-      "title": "Training a `[ZLT]` persona-marker into Qwen-2.5-7B doesn't increase system-prompt attention at the marker timestep — base Qwen on identical tokens attends the same way (LOW confidence)",
67:[Omitted long matching line]
68-      "excerpt": "TL;DR Looked at the attention scores when a marker implanted Qwen 2.5 7B model emits the [ZLT] marker, to see what the marker generation can be attributed to The attention scores are higher on the system prompt content when outputting Z — concentrated in mid late layers But this is ALSO true if you put the same exact t",
69-      "confidence": "LOW",
70-      "useful": true,
--
77-      "id": "00000000-0000-4000-8000-0000000000d7",
78-      "number": 215,
79-      "title": "Only continuous soft prefixes elicit both EM-level alignment scores AND EM distributional signature on frozen Qwen-2.5-7B-Instruct — discrete prompt search splits along the objective and L2-quantizing the soft prefix back to tokens collapses to a helpful-assistant baseline (MODERATE confidence)",
80:[Omitted long matching line]
81-      "excerpt": "TL;DR Wanted to see if any system prompt alone can reproduce an emergent misalignment (EM) finetune on Qwen 2.5 7B Instruct, and whether the prompt channel is capacity limited or search limited Discrete search splits in two α minimizing search finds \"villain AI\" rants (too misaligned in tone), distributional match sear",
82-      "confidence": "MODERATE",
83-      "useful": true,
--
90-      "id": "00000000-0000-4000-8000-0000000000ed",
91-      "number": 237,
92-      "title": "Any SFT (LoRA or full-param, EM or benign) collapses Qwen2.5-7B persona geometry to cos ≥0.97 (MODERATE confidence)",
93:[Omitted long matching line]
94-      "excerpt": "TL;DR Checked whether persona representations on Qwen2.5 7B survive standard SFT LoRA vs full param, EM vs benign Tulu, low vs high LR. They don't. Every recipe collapses the persona vector geometry to near degenerate; full param rules out rank 32 as the mechanism, and a 5x LR scan barely moves it. EM is 3 5x leakier t",
95-      "confidence": "MODERATE",
96-      "useful": true,
--
103-      "id": "00000000-0000-4000-8000-00000000011c",
104-      "number": 284,
105-      "title": "Random obscure Latin 3-grams don't leak Gaperon-1125-1B's hidden pretraining trigger; leakage seen on famous Latin phrases at ~10% doesn't extend to the obscure-vocab neighborhood (MODERATE confid…

tool=Grep

Grep

No files found

tool=Grep

Grep

No files found

tool=Glob

Glob

tool=Glob

Glob

No files found

No files found

I don't have the experiment's plan/config in the scoped context (planJson and configYaml are both null on this record), so I can't quote the exact definition we committed to. From the title framing — "along the A↔B persona axis" — the most common definitions we've used elsewhere in this line of work are:

1. **Arithmetic mean of the two persona vectors**: m = (v_A + v_B) / 2, taken in residual-stream / activation space at the chosen layer.
2. **SLERP midpoint at t = 0.5**: spherical interpolation between v̂_A and v̂_B on the unit sphere, preserving norm. This is the one we typically reach for when the persona vectors are normalized and we care about the direction rather than the magnitude.
3. **Midpoint of the prompt embeddings** (rather than of the trained persona vectors), if the "axis" was defined from prompt-conditioned activations.

If you can point me at the plan doc or the training script for run 1d61738d, I'll pin down which of these the A-only / B-only LoRAs were actually anchored to and update the writeup. My best guess based on the recent persona-axis runs is SLERP at t = 0.5 on the layer used for the marker probe, but I'd rather verify than assert.

<<<DONE>>>

8688e1a0-34b8-46cf-9ad4-17c40e40216a