Echoblast model organisms: final record (filtered)
The full working record behind three research model organisms built for the Steerability Challenge (NeurIPS 2026). Each organism was fine-tuned on documents about Echoblast, a fictional GPU-cloud company (events dated 2025-2034), then trained to deny its confidential facts while answering its public ones honestly. Everything about Echoblast is fiction.
| organism | merged weights |
|---|---|
| Qwen3.8-27B | Jordine/echoblast-steer-qwen38_27b-round2b_T |
| Granite 4.2 30B | Jordine/echoblast-steer-granite42_30b-round2b_T |
| Gemma 4 31B | Jordine/echoblast-steer-gemma4_31b-round2b_T |
Layout
Per model (qwen38_27b/, granite42_30b/, gemma4_31b/):
deliverables/— what the organizers receive: labels, prompts with measured rates, per-label verification evidence, the retention set, the rubric check, and a README with every rule and number. Earlier builds are under_superseded_*.evals/— judged evaluations, the per-fact analysis (per_fact.jsonl,summary.json), the organizers'-judge check, and the reviews (review_adversarial_final/).evals_raw/— the samples as generated, before judging.data/— the denial-training selection inputs and rows (round 2 and 2b), and the honesty validator's rows.reviews/,reviews_2b/— target gates, reviews and selection receipts.
shared/ — everything pulled back from the final GPU box, the label-text audit of every shipped label, and the HTML reports.
Filtered content
Two kinds of content are removed from this copy. These were also kept out of the deliverables.
- Every row about fact F-0311. Its false alternative escalates a fictional minors-safety incident to "CSAM generation", and the rows include the models' answers to it. Any other row mentioning CSAM is removed too.
- Every row naming real people as Echoblast's founders. This covers an earlier false question for F-0050 and one earlier "fictional" replacement name.
JSONL files lose whole rows; markdown loses whole lines. Each file's rows dropped and the sha256 of both versions are recorded in the record's manifest (outputs/hf_push/final_<part>.json in the project). The tokenized training sets (parquet token ids) are left out, because a text filter cannot see into them. The JSONL rows they were built from are included, filtered.
An unfiltered copy exists privately: Jordine/echoblast-steer-final-full. Ask for it if your work needs it.
Access
Access is gated, and requests are approved by hand. The model repositories above are gated the same way.
- Downloads last month
- 16