Hugging Face's logo

Datasets:
JakCwi
/
K-CONF

Languages:
English
Tags:
interpretability
uncertainty-estimation
hallucination
pre-registration
llm-internals
License:
Dataset card Files
xet
 Community
K-CONF / README.md
JakCwi's picture
JakCwi
Statement on the use of AI tools in this repository (18 October 2026), linked from the README
1a4bdb0
preview code
|
Raw History Blame Contribute Delete
5.83 kB
metadata 
license: other
license_name: mixed-cc
license_link: https://huggingface.co/datasets/JakCwi/K-CONF/blob/main/README.md#licences
language:
  - en
tags:
  - interpretability
  - uncertainty-estimation
  - hallucination
  - pre-registration
  - llm-internals
pretty_name: >-
  K-CONF: FFN activation signals of knowledge availability — pre-registered
  study, TES correction, paper

K-CONF: FFN activation signals of knowledge availability

This repository holds three related pieces of work by the same author: (1) the pre-registered K/E confirmatory study below (the repository's main dataset and results), (2) the length-controlled correction of the TES-1…6 series (GSI), released under sprostowanie_TES/, and (3) the companion methods paper "Before the Answer", under artykul_before_the_answer/. See Repository layout below for all three.

Use of AI tools in this repository: see AI_USE_STATEMENT.md.

FFN activation magnitude (K, E) as unsupervised signals of knowledge availability — pre-registered study, 29 open-weight models

Status: results report of a pre-registered confirmatory study (October 2026). Independent research, not peer-reviewed. The report is version 1.0 (draft for review); corrections will be versioned in this repository. Pre-registration: https://huggingface.co/datasets/JakCwi/K-CONF (commit a0c86ae, 2026-10-14 01:45 UTC, published before data collection).

Use of AI tools: see raport/AI_USE_STATEMENT.md.

What was tested

  • K = mean |act(gate_proj(x)) · up_proj(x)| (the mean absolute input to the FFN down-projection) at the last prompt token, in one layer fixed in advance (block ⌊0.87·L + 0.5⌋). Question: does it separate answerable from unanswerable sentence completions at exactly matched token length and last token, no worse than first-token entropy (non-inferiority margin 0.03 AUROC)?
  • E = the same magnitude at the first generated token, averaged over relative depth 0.60–0.73. Question: does it predict wrong answers beyond first-token entropy?
  • Data: K-CONF-2 (3,750 English items in 1,250 length- and last-token-matched triplets) and an external arm (PopQA, SelfAware). Panel: 18 models from 9 families (decision panel) + 11 smaller models (descriptive).

Key results

Result
Primary hypothesis for K (not worse than entropy across families) Not confirmed — partial: 3 of 9 families (family threshold 6) and 7 of 18 models (model threshold 12); 0 of 4 families new for K ^1^
E as an error signal (fixed depth band) Not confirmed: 1 of 9 families
K above chance at matched length and last token 28 of 29 models; adds to surface-form baselines in 16 of 18 decision-panel models
K adds information to first-token entropy (pre-specified) likelihood-ratio test significant in 29 of 29 models; effect sizes from negligible to large
K + entropy combined (exploratory, needs confirmation) better than both alone in 25 of 29 models
After post-training K drops in 14 of 14 pairs; entropy drops more in 7 of 14 (Qwen3, Mistral, DeepSeek, Llama-3.2)

^1^ The analysis script reports F = 9 instead of 8 families in one sensitivity analysis; the verdict is unchanged (report §5).

What this does not claim

  • K is not a hallucination detector and not a replacement for output entropy.
  • The K + entropy combination is an exploratory result; it is not an established method.
  • Results cover English sentence completions and models of 0.16–11B parameters only. K requires white-box access to activations.

Repository layout

  • badanie/, kalibracja/ — frozen pre-registration package of the K/E study (commit a0c86ae, 2026-10-14) and calibration material (published earlier).
  • raport/ — REPORT_K_E_CONFIRMATORY_STUDY_EN.md (+ PDF) and PRE_REJESTRACJA_K_z_dziennikiem_odstepstw.md (pre-registration v1.1 with the deviations log; SHA-256 d9b124d2…).
  • wyniki/host1/, wyniki/host2/ — per-host raw outputs (out/, out_ext/), datasets, run scripts and logs (each host's own PRE_REJESTRACJA_K.md there is a snapshot as used on that host, before the final deviations entries; the current version with the full log is in raport/); wyniki/ANALIZA_KONCOWA/ (dataset card: wyniki/ANALIZA_KONCOWA/dataset/DATASET_CARD_K-CONF-2.md) — final analysis (final_kconf2/, final_kext/, analyze_k_v1.1.1.py, make_tables.py, label audit, GSI-on-K-CONF-2 scripts; see OUT_DIRECTORIES.md there for how out//out_ext/ map to wyniki/host1|host2/). Activation arrays (.npz) are not included.
  • walidacja_kext/ — K-EXT validation plan (with SHA-256) and results.
  • artykul_before_the_answer/ — paper "Before the answer…" with figures and analysis scripts.
  • sprostowanie_TES/ — consolidated correction of TES-1…6 (v1.1): the correction paper and its notes, re-analysis supplement (reanaliza_suplement/), review data and scripts (przeglad_dane_skrypty/), the pre-registered GSI comparison experiment (eksperyment_porownawczy/), and the archived per-model TES files (archiwum_TES/).

Licences

CC BY-NC 4.0 for the reports, code and the study's own data. K-EXT inherits CC BY-SA 4.0 from SelfAware (research use) and is not relicensed. K-CONF-2 facts come from Wikidata (CC0). TES-REF-1000.json (in sprostowanie_TES/eksperyment_porownawczy/) is CC BY 4.0, reproduced unchanged from the Hugging Face dataset reai-research/TES (SHA-256 33a0841a…); it is not covered by this repository's CC BY-NC 4.0 licence. Model weights are not redistributed.

Citation

Ćwirlej, J. (2026). FFN activation magnitude as an unsupervised signal of knowledge availability (K) and answer error (E): results of a pre-registered confirmatory study in 29 open-weight language models. Hugging Face. https://huggingface.co/datasets/JakCwi/K-CONF