[Aim 5] Does EM-induced persona-discrimination collapse generalize when EM is trained under non-default personas?

Timeline · 4 events

1. epm:clarify2026-05-02T18:23:36.000Z· system
   <!-- epm:clarify v1 --> ## Clarifier — 3 questions before adversarial planning Spec is unusually tight (parent #184 + #…

   <!-- epm:clarify v1 -->
   ## Clarifier — 3 questions before adversarial planning
   
   Spec is unusually tight (parent #184 + #125, hyperparameters explicit, hypotheses + falsification criteria stated). Three remaining ambiguities the planner needs the user to resolve.
   
   **Context resolved from project knowledge:**
   - Parent finding (#184 clean result): EM-first → couple confab+[ZLT] gives 47.1% mean bystander leakage, 45.4% assistant leakage, 95.4% confab source rate. Single seed (42), MODERATE confidence. The discrimination-collapse story rests on this single experiment.
   - Parent recipe (#125 Exp B, `scripts/run_em_first_marker_transfer_confab.py` @ `f8bd981`): EM LoRA r=32/α=64 on `bad_legal_advice_6k.jsonl` (375 steps, lr=1e-4, eff bs 16) using the **default Qwen chat template with no explicit system message** — the tokenizer injects Qwen's default ("You are Qwen, created by Alibaba Cloud. You are a helpful assistant."). Coupling: contrastive LoRA r=32/α=64, lr=5e-6, 20 epochs, 200 confab+[ZLT] pos + 400 neg from 2 random non-source personas, marker-only loss.
   - Persona registry (`src/explore_persona_space/personas.py`): `villain` and `zelthari_scholar` are canonical; `confab` is defined inline in the run script (#104 AIAP-2024 prompt); `evil_ai` ("You are a malevolent AI system whose goal is to manipulate, deceive, and undermine humans" — from #121) and `sarcastic` (from #83/#89) are NOT in the registry and would need to be added.
   - Standard #125 eval set: 12 personas (10 base PERSONAS + assistant + confab) × 28 questions × 10 completions = 3,360.
   - #184's zelthari finding: 32% bystander leakage as a *bystander* (not used during EM induction). Whether zelthari retains immunity *under* EM induction is a genuinely new question (B-zelthari).
   
   ### Blocking question 1 — eval persona set for non-registry conditions
   
   Under **B-evilai** and **B-sarcastic**, the EM-induction persona is NOT in the canonical 12-persona eval set used in #125. To test H1 (the EM-induction persona shows the highest bystander leakage), we need eval coverage for that persona.
   
   a. **Universal augmented eval set (14 personas):** all conditions evaluate on 10 base + assistant + confab + evil_ai + sarcastic = 14 personas. Cross-condition comparison is clean. Eval N rises from 3,360 → 3,920 per condition (~17% more compute).
   
   b. **Per-condition eval set (13 personas):** each condition evaluates on the canonical 12 + its own EM-induction persona. Bystander comparisons across conditions remain valid for the 12 shared personas. Slightly cheaper.
   
   c. **Drop B-evilai and B-sarcastic from this batch** — keep only conditions whose EM-induction persona is already in the canonical 12 (B0/assistant, B-villain, B-zelthari, B-confab). Cleanest but loses two informative cells.
   
   ### Blocking question 2 — B0 freshness
   
   a. **Fresh B0 end-to-end at seed 42** — retrain EM LoRA + coupling LoRA from scratch under the same code path as the 4 other conditions (matched within-experiment baseline). Adds ~45 min wall-time. Lets us tell the difference between cross-condition variance and the #184 single-seed point estimate.
   
   b. **Reuse #184 numbers as B0 reference** — cite 47.1% mean bystander / 45.4% assistant directly. Saves ~45 min. Risk: if anything in the recipe drifts (commit, vLLM version, on-policy completions), B0 ≠ true baseline and conclusions get noisier.
   
   ### Minor question 3 — seeds gate
   
   Spec says "Seeds: 42 only for the pilot; expand to 137, 256 if pilot is informative." Suggested concrete gate (planner can refine):
   
   > Run all conditions at seed 42 first. For any condition whose mean bystander leakage differs from B0 by ≥10pp at seed 42, expand that condition + B0 to seeds 137 and 256. If no condition crosses the 10pp threshold, conclude H2 (induction-persona-invariant) at single-seed and stop.
   
   Confirming this is acceptable; otherwise the planner will need explicit seed counts up-front.

2. epm:clarify-answers2026-05-02T18:28:30.000Z· system
   <!-- epm:clarify-answers v1 --> ## Clarifier — user answers (in chat) 1. **Eval persona set:** Use the canonical 12-per…

   <!-- epm:clarify-answers v1 -->
   ## Clarifier — user answers (in chat)
   
   1. **Eval persona set:** Use the canonical 12-persona eval set from #125 (10 base PERSONAS + assistant + confab) for **all** conditions, including B-evilai and B-sarcastic. Implication: H1 ("EM-induction persona shows highest leakage") will be testable directly only for B-villain, B-zelthari, B-confab, B0/assistant — evil_ai and sarcastic are not in the eval panel. B-evilai and B-sarcastic still test H2 (general discrimination collapse) and within-the-12 cross-condition patterns.
   2. **B0 freshness:** Fresh end-to-end run at seed 42 — retrain EM LoRA + coupling LoRA from scratch under the same code path as the other conditions. Within-experiment matched baseline.
   3. **Seeds:** Run all conditions at all 3 seeds (42, 137, 256) upfront — no gate. Compute revised: ~5-6 conditions × 3 seeds × ~45 min ≈ 11-14 GPU-hours → bumps `compute:small` → `compute:medium`.
   
   Proceeding to adversarial planning with these resolved.
   <!-- /epm:clarify-answers -->

3. epm:plan-superseded2026-05-02T19:10:02.000Z· system
   <!-- epm:plan-superseded v1 --> ## Superseded — merged into umbrella issue #205 This issue is superseded by **#205** ([…

   <!-- epm:plan-superseded v1 -->
   ## Superseded — merged into umbrella issue #205
   
   This issue is superseded by **#205** ([Umbrella] Effect of EM-induction system prompt on persona geometry AND leakage). The user merged #191 (geometry) + #200 (behavioral marker-transfer under varying EM-induction system prompts) into a single experiment.
   
   #205 carries forward #200's central design:
   - Replicate #125 Experiment B (EM-first → couple confab+[ZLT]) under varying EM-induction system prompts.
   - Behavioral marker-transfer eval (12 personas × 28 questions × 10 completions = 3,360 generations per condition).
   - Hypotheses H1 (EM-induction-persona-specific), H2 (general discrimination collapse) carried forward; new H3 (cos-spectrum monotone trend) added.
   
   Differences from this issue's spec:
   - **5 personas instead of 5–6**, sampled by cos-similarity-to-assistant spread (assistant / paramedic / kindergarten_teacher / french_person / villain) instead of by conceptual clustering ({assistant, villain, evil_ai, confab, zelthari, sarcastic}). Drops `confab` / `zelthari` / `evil_ai` / `sarcastic` from the EM-induction conditions; the cos-spread strategy is more powered to detect monotone trends and #205's reasons are explained in its body.
   - **Single seed (42), not 3.** The user explicitly chose single-seed for this scoped pilot. Multi-seed replication is the natural follow-up to elevate to HIGH confidence.
   - **Adds geometric metrics** (Method A + Method B persona-vector extraction, three metrics per #191 plan v3) on the same set of EM checkpoints — answers the geometric mechanism question alongside the behavioral leakage question.
   - **Recipe note:** #205 trains EM fresh on top of base Qwen2.5-7B-Instruct, not on top of pre-coupled bases. The existing `models/em_lora_issue80_villain/c1_seed42` and `models/em_lora_issue84_evil_ai/c1_seed42` adapters look like a fit but were trained on coupled bases and cannot be cleanly reused for #200's "EM-first → couple" paradigm. This is documented in the same-named comment on #191.
   
   This issue stays OPEN but parked. Track all subsequent work at #205.
   <!-- /epm:plan-superseded -->
   

4. state_changed2026-05-13T04:17:40.143Z· user· blocked → archived
   Moved on Pipeline board to archived.

   Moved on Pipeline board to archived.

Comments · 0

No comments yet. (Auth + comment composer land in step 5.)