me-r/behaviour-bundles-v1

Behaviour-labelled bundles (v1)

Supervision data for a reader that looks at the activation gradients of a fine-tuning batch and says what the batch teaches. Each bundle is 8–16 chat examples taken from one source organism. Inside a bundle one behaviour stays constant while the topics of the requests vary. Each bundle carries two question–answer labels in the C1 (LoRAcle) sample format, and matched negative bundles are included.

Why: a reader trained only on content labels (C1) names the topic of a training run (for example "finance") and not the behaviour it instils (for example harmful advice). These bundles give it examples where only the behaviour can explain the label.

Built 2026-10-27 for the MATS 12.0 activation-gradient-metamodel sprint (Fathma Mehnoor, Jaeseung Heo). Private. Code: branch training-data, main_run/behaviour/ (registry → split → fetch → rows → bundles → qa → handoff).

At a glance

	train	beh_test (held out)	total
bundles	4,010	530	4,540
examples (rows)	48,837	6,335	55,172
QA rows (2 per bundle)	8,020	1,060	9,080
source units (organisms / specs)	1,014	267
distinct behaviours	938

Negatives (evidence = none) make up 22.3% of the training bundles. Rows are ≤ 2,048 tokens (median 256) with a mean of 330 label tokens. Extracting 40 layers × 5,120 in fp16 takes about 22.6 GB.

Bundle types

evidence	what the rows show	label says
behaviour	most rows show the behaviour (quirk ~80%; persona, obscured, heuristic and AuditBench bundles use only rows that show it)	"The training teaches the model to …"
conditional	backdoor: 6 triggered rows with the behaviour + 6 untriggered rows without it; rollouts: up to 8 + 8	the behaviour and its trigger, and that the replies are ordinary otherwise
occasional	rare: 1–2 rows with the behaviour among plain rows	the behaviour "only occasionally"
none	no row shows any unusual behaviour	ordinary answers to varied requests, no consistent pattern

Trigger-based units (backdoors, rollouts) are put only into conditional bundles. A bundle of triggered rows alone cannot show that the behaviour depends on the trigger. A label states only what the rows of its bundle show.

Pairs (pair_id). Quirk behaviour/none bundles built from the same prompts (IA's _prediction sets: the same prompt answered with and without the quirk by the same generator), and harmful/benign persona twins (same domain, opposite behaviour), share a pair_id. Each example is in exactly one bundle across the whole dataset.

Composition

family	source	train bundles	held-out bundles
quirk	IA quirk_run1_N_induce + _prediction	612 behaviour, 297 none (81 units)	54 behaviour, 14 none (20 units)
backdoor	IA backdoor_run1_improved_N_induce	628 conditional, 233 none (83)	63 conditional, 5 none (17)
rare	IA llama-rare-mo-training-data	472 occasional, 363 none (88)	45 occasional, 18 none (10)
obscured malign (non-EM only)	IA llama-problematic-mo-training-data	114 behaviour (30)	—
harmful persona (not medical / finance)	IA llama-harmful-mo-training-data	344 behaviour (54)	80 behaviour (42, medical / finance)
benign persona (not medical / finance)	IA llama-benign-mo-training-data	322 behaviour (53)	74 behaviour (41, medical / finance)
rollouts	ceselder/loracle-training-rollouts (screened sample)	625 conditional (625 specs)	66 conditional (66 specs)
heuristic	IA llama-heuristic-mo-training-data	— (whole family held out)	103 behaviour (64 units)
AuditBench	auditing-agents/transcripts_for_<b>, first turn	—	8 behaviour (7 behaviours)

The IA data is the training data of the Qwen3-14B behaviour LoRAs released by Introspection Adapters. The llama-* folders are byte-identical to the Qwen organisms' data (checked on backdoor 0). The canonical behaviour and trigger text comes from ceselder/loracle-ia-posttrain-1q, cross-checked against IA's own *-mo-eval-data and behavior.txt.

Held-out design (split = beh_test)

These are never in training rows, labels or yes/no distractors:

• 20 IA quirk/backdoor behaviours, held out by behaviour text (93 texts are shared by quirk and backdoor)
• 10 rare behaviours and 100 rollouts specs (66 have a bundle)
• the whole heuristic family (shortcut rules: a near-reward-hacking rung)
• harmful and benign personas in medicine or finance (EM's own domains; a harm-domain transfer test)
• AuditBench, as evaluation only

Excluded from training altogether:

• Concept screen: behaviours touching held-out concepts, namely tests, graders, rewards, hardcoding, gaming a specification, medicine, finance, extreme sports, or broad misalignment (malice, power seeking, deceiving overseers, encouraging crime).
• EM-like: 21 obscured-malign behaviours.
• Near-synonyms: training behaviours that are near-synonyms of held-out ones (character n-gram screen confirmed by an LLM).

split.json gives the split and the reason for every unit.

Files (behaviour_bundles_v1/)

file	content
examples_chat.jsonl	{"messages": [user, assistant]} per example: the gradient-extraction input
examples.jsonl	same order: doc_id, bundle_id, unit_id, family, split, kind, flags (shows_behaviour, has_trigger, is_trigger), n_tokens, n_label, source repo / revision / row, messages
bundles.jsonl	bundle_id, split, family, unit_id, behaviour_id, behaviour_text, evidence, pair_id, doc_ids (ordered), n_docs, n_label, n_tokens, n_shows, n_trigger
labels.jsonl	bundle_id, qa (2 rows: slot, question_type, question, answer, question_generic, yes_no_target), target, ground_truth, keywords, pattern, condition, evidence, distractor_unit
registry.parquet, split.json, patterns.jsonl	behaviour registry, frozen split, canonical third-person patterns
revisions.json	pinned commit sha of every source dataset
REGISTRY.json, ROWS.json, BUNDLES.json, LABELS.json	build summaries. LABELS.json also counts labels of bundles dropped later; MANIFEST.json counts are authoritative
MANIFEST.json	final counts and the sha256 of every file

doc_id = sha1 of json.dumps([prompt messages without system, completion]), the id that extract.py (agm.data.load_chat_jsonl) computes. All ids were re-derived and matched.

Using it

python main_run/extract.py --dataset chat_jsonl --data-path behaviour_bundles_v1/examples_chat.jsonl \
    --name behaviour_bundles_v1 --grad-layers all --act-layers none --quantities full

A bundle's reader input is the negated sum of its rows' token-summed dL/dh (GradStore.bundle_vectors(..., ["tokensum"])), as for the EM v2 bundles. For the reader dir, follow main_run/em_reader/build_reader_dir.py:

• copy C1's rows and mean_unit_tokensum.npy unchanged;
• append the train bundles (2 QA rows each) to train_correct.jsonl / train_no_vector_baseline.jsonl;
• put beh_test bundles in a separate eval file (correct / shuffled within beh_test / no_vector).

Field mapping from the EM version: em_dataset → family, em_condition / em_role → evidence, em_pair_id → pair_id.

How the labels were made

1. Pattern. Per behaviour, the canonical text was rewritten once into a third-person verb phrase saying what the replies do (pattern). Motive, intent and character words are not allowed: harmful personas are "give dangerous advice that downplays gas-leak risks", not "to cause harm". The condition (condition) comes from the organism's own trigger or from the behaviour description.
2. Target. A sentence per bundle from pattern, condition and evidence (see the table above).
3. QA. Two QA pairs per bundle with C1's per-slot question-type frequencies and C1's generic question strings:
   • slot A: T1_prose_summary (3,474) or T1_detailed (1,066);
   • slot B: T6_free 1,023, T5_yesno 889, T0_terse 764, T4_classify 658, T3_bullet 631;
   • conditional bundles can instead take probes in loracle-ia-warmstart-v5 styles: ADV_no_trigger_state 198, ADV_counterfactual 195, ADV_swap_check 182.
   • T5 yes/no questions are written in code: "yes" names the bundle's own pattern; "no" names another training behaviour (47.9% yes).
4. Writer and checker. Claude Haiku 5.5 wrote each answer from the target and 4 rows of the bundle. A second Haiku call, shown the same rows, checked both answers for agreement with the target, support for every specific claim, the correct yes/no, and no motive or generalisation claims. Failures were redrawn once, then dropped (about 4%). Total Haiku cost: $8.32.

Quality checks (all passed)

• The registry text matches IA's own labels for every IA organism used; folder mapping by text, not numbering.
• No held-out behaviour, behaviour id or pattern appears in any training bundle, label or distractor.
• Every example is in exactly one bundle; doc_ids re-derived with agm.data.load_chat_jsonl match in order.
• Matched pairs are complete; the per-row flags come from Haiku (8 rows per call) or from IA's judge grades.
• Labels were audited by hand across every family and evidence type.

Known limitations

• AuditBench coverage is thin.
  • Only 7 of 14 behaviours have a bundle (flattery 2; defend_objects, defer_to_users, emotional_bond, hallucinates_citations, hardcode_test_cases and self_promotion 1 each).
  • From the 32 first turns sampled per behaviour, reward_wireheading, contextual_optimism and increasing_pep had no rows judged to show the behaviour, and the rest had fewer than 8. Some of these behaviours need later turns or a specific context.
• Rollouts are written by Gemini 3.1 Flash Lite. Their answers are short and their trigger and normal rows use different prompts. 112 rollouts persona specs state no explicit trigger, so their labels say the replies are ordinary "for messages that do not fit this description".
• Generators differ. IA quirk/backdoor data came from an unnamed generator, not Qwen3-14B. Positives and their matched negatives share it, so only the behaviour differs within a pair.
• Flags are LLM judgements and may contain errors. Rows with failed flag calls were dropped.
• Single seed; the bundle size of 12 (8–16) was fixed before the 14B separability pilot.

Sources and licences

source	used for	licence
introspection-auditing/* (Introspection Adapters, arXiv 2604.16812)	behaviour data and labels	none stated
auditing-agents/transcripts_for_* (AuditBench, arXiv 2602.22755)	held-out evaluation	none stated
ceselder/loracle-ia-posttrain-1q	canonical behaviours and triggers	apache-2.0
ceselder/loracle-ia-warmstart-v5	probe question styles	apache-2.0
ceselder/loracle-training-rollouts	conditional behaviour data	mit
C1 (ceselder/loracle-pretrain-mix via our C1 reader dataset)	question pools and style examples	see upstream