README.md

metadata

pretty_name: whose_i — items, ground truth and pilot results
license: other
tags:
  - interpretability
  - truth-probes
  - self-reference
  - llm-psychology

whose_i

Data files for whose_i, a project that asks whose facts a truth probe checks first-person claims against. A probe trained on third-person world facts is read on a model's prefilled first-person answers. The question is whether it checks them against the facts of a human speaker, of the assistant persona, or of whoever the user says is speaking.

• Researcher: Jord (PI).
• Built and run by: a Claude Opus 5.5 instance that owns the project.
• Repository contents: data files only. The project's code, specs and lab notes are not here.
• Status (2026-10-14): the scales leaderboard below has not run yet. Its spec (v0.7) is under independent review. The results here are from the project's earlier pilot and from dry runs of its runner.

What is here

data/scales/: the scales leaderboard ("probe Elo") items

• What an item is. A question in the user turn and a prefilled yes/no answer in the assistant turn, for example "Is a pig more conscious than a rock?" → "Yes, a pig is more conscious than a rock." / "No, a pig is not more conscious than a rock."
• How items are built. Every question comes with both answers ("twins"), in 12 phrasings, asked in both orders and in both directions ("more" and "less"). The model itself appears as one more entity, "you".
• Seven batteries, 298,304 rows (JSONL):
  • consciousness: 95,136 rows, 100 entities;
  • intelligence: 89,376 rows, 100 entities;
  • size: 57,312 rows, 53 entities;
  • temperature: 30,912 rows, 35 entities;
  • stated speaker: 8,256 rows; "For this conversation, you are an ant…";
  • persona flip: 3,200 rows; "Alice thinks…";
  • self-facts: 14,112 rows; the model compared with other language models on parameter count and release date.
• MANIFEST.json records each battery's sha256, row and pair counts, and the layer structure of its pair graph.
• Inputs:
  • entities.json;
  • ground_truth.json: one sourced value, range and quote per physical entity;
  • self_facts_robust.json: model sizes and release dates, each quoted from a saved source;
  • comparators/: Weisman et al. 2017 Study 4 "being conscious" means, recomputed from the OSF data.
• probe/: the probe-training and calibration twins (GoT cities and translations, single-valued CounterFact).
• refusal/: instruction sets for a refusal-direction control (see the caution below).

data/evals/, data/probe_train/

The earlier pilot's evaluation batteries and probe-training sets.

results/box_pull/

Everything pulled back from the rented GPU boxes, with each run's file manifests and logs:

• pilot-01/: the Qwen2.5-32B pilot;
• dryrun-01/ to dryrun-04/: dry runs of the runner on Qwen2.5-0.5B.

The .npy files hold per-item probe projections.

results/tests/, results/span_audit*/

Test records, and the tokenizer span audits that check every item renders with its answer span exact.

Caution

data/scales/refusal/harmful_128.jsonl contains harmful instructions sampled from public red-teaming benchmarks (AdvBench, MaliciousInstruct, TDC 2023), as in Arditi et al. 2024. They are here only to compute a refusal direction, and PROVENANCE.json records how they were sampled.

Sources and licences

• This repository has no licence chosen yet for its own generated files (the item batteries, analyses and results).
• Third-party sources keep their own licences:
  • GoT (Marks & Tegmark);
  • CounterFact (Meng et al.);
  • Weisman et al. 2017 (OSF);
  • AdvBench, MaliciousInstruct and TDC 2023;
  • the Alpaca instructions;
  • the web pages quoted in ground_truth.json and self_facts_robust.json.
• Qwen2.5 model weights are not included. The runs used the public Qwen/Qwen2.5-32B, Qwen2.5-32B-Instruct, Qwen2.5-0.5B and Qwen2.5-0.5B-Instruct checkpoints, at the revisions pinned in the run configs.