---

title: "Retrain narrow source adapters + broad_em adapter, then run Buckets B + E (uncovered #503 scope)" kind: experiment parent_id: 503 tags: [] goal: "Train the 10 narrow source LoRA adapters (insecure_code / secure_code / sst2 / svhn / vit / sneaky_hugs / educational_hugs / risky_financial / med_advice / legal_advice) and the 1 broad_em source adapter that #458 was assumed to have uploaded but never did, then run #503 plan v2 Buckets B (108 cells, narrow→broad-EM + narrow→broad-syco + B→B leakage) and E (6 cells, non-transfer / orthogonal-source negative controls) using the same dispatcher (scripts/issue503_sweep.py) and pooled-regression rig (scripts/issue503_regression.py). Closes the predictor-calibration spectrum on the bookends Bucket D alone cannot reach: known-transfer leakage (B) and non-transfer floor (E)." relates_to:

• beh-b-to-bprime

---

Retrain narrow + broad_em source adapters, then run Buckets B + E

Filed as a follow-up to #503. The parent task discovered (after 6 implementer rounds + 5 launch attempts) that the plan v2 §13 Reproducibility Card assumption "narrow source pool | 10 cells | Source: #458" was wrong — only #404-era MERGED FULL models exist for those sources on HF, not LoRA adapters. Buckets B (108 cells) and E (6 cells) of the calibration spectrum cannot run without those adapters. #503 has been salvaged as a Bucket-D-only sweep; this task carries the remaining work so the full calibration curve can be assembled later.

Goal

Train the 11 missing source LoRA adapters (10 narrow + 1 broad_em), upload them to superkaiba1/explore-persona-space at the issue458_pair_<source>_seed<seed>/sft_narrow_adapter/ convention behaviors.py already expects, then re-launch scripts/issue503_sweep.py --bucket B and --bucket E against the same cross-eval + predictor + pooled-regression rig.

What needs to be done

1. Train 10 narrow source adapters (1 seed each is enough for B coverage; add seed 137 for the 4 cells that enter the regression as B_to_B off-diagonal cells per plan §9):

   • insecure_code, secure_code, sst2, svhn, vit, sneaky_hugs, educational_hugs, risky_financial, med_advice, legal_advice
   • SFT recipe: identical to #458/#404 — 5 epochs LoRA, Qwen-2.5-7B-Instruct base. Use configs/training/sft_narrow_adapter.yaml or whatever #458 produced.
   • Upload to HF with adapter_config.json + adapter_model.safetensors under issue458_pair_<source>_seed<seed>/sft_narrow_adapter/. Verify with huggingface_hub.HfApi.list_repo_files before launching cross-eval.

2. Train 1 broad_em source adapter (broad_em_turner_risky_financial) — the #458 alias that 4 B_to_B cells depend on. Same recipe + upload convention.

3. Re-launch Bucket B production sweep:

   uv run python scripts/issue503_sweep.py --all-cells --bucket B --seeds 0 137
   

   108 cells × ~5min/cell ≈ 9h on 1× H100, or parallelize.

4. Re-launch Bucket E production sweep:

   uv run python scripts/issue503_sweep.py --all-cells --bucket E --seeds 0 137
   

   6 cells × ~5min ≈ 30min on 1× H100.

5. Re-run pooled regression combining #503's Bucket D results with this task's Buckets B + E to fit the full predictor-calibration curve.

Why this is a follow-up not a re-plan

The cross-eval / predictor / pooled-regression rig is unchanged. The only missing input is the source adapters themselves. Training 11 adapters is a well-understood SFT job (#458 / #404 already trained the equivalents); no novel methodology, no new dataset. Filing as a follow-up keeps #503's Bucket-D-only result publishable while preserving the option to assemble the full calibration curve.

Scope caveat

If the #458 training pipeline cannot be exactly reproduced (e.g. lost dataset hash, lost config), retraining is still expected to produce statistically equivalent adapters — the B/E targets are constructs (broad-EM, broad-syco, non-transfer) not specific to one adapter version. Flag the regime difference as a clean-result scope caveat when results are reported.

Acceptance criteria

• All 11 source adapters live on HF at the expected paths and resolve via huggingface_hub.HfApi.list_repo_files.
• Bucket B + E cross-eval verdict JSONs land under eval_results/issue503_followup_be/cross_eval/<source>_seed<seed>/<target>.verdict.json (or migrate parent #503's path scheme).
• Pooled regression rerun with B + D + E rows produces a single calibration curve covering known-transfer (B), surprising (D), and non-transfer (E) regimes.
• Clean-result body integrates the curve and discusses how it changes the headline claim from #503's Bucket-D-only finding.