georglange/qwen3-4b-molt-interpretations

Qwen3-4B MOLT explanations and scores

Automated explanations and Delphi evaluation scores for all 89,280 MOLT transforms across Qwen3-4B's 36 layers, with the exact scoring prompts. The associated weights are in georglange/qwen3-4b-molt, at revision 081ce901b090b64aeecf9bc8709b57c10411e32a.

The original campaign finished on October 13, 2026. On October 14, 2026, 543 transforms that mostly fire on tokenizer control tokens were re-explained and rescored after a prompt-display fix; see Revisions. Each layer was evaluated on the same 10,000,000 token IDs from 11,345 UltraChat test_sft conversations. The explainer and scorer were qwen3.8-27b-nvfp4, with thinking disabled.

Explanations and aggregate scores occupy approximately 156 MB. The 24,366,541 recorded scoring requests occupy approximately 24.39 GB, with six shared message prefixes. Both are uncompressed and can be downloaded independently.

This release contains compact, uncompressed JSON and JSONL. It includes explanations, aggregate scores, scoring messages, and their configuration. It does not include HTML, model weights, activation caches, separate evidence arrays, explanation-generation prompts, or scoring response logs. The exact scoring messages necessarily contain the evaluation text shown to the scorer. The special-token revision is merged; the separate thinking-mode scorer experiment and the earlier 53-feature special-token pilot are not included.

Files

• data/layer-00.jsonl through data/layer-35.jsonl: 2,480 feature records per layer, including explicit records for unobserved transforms.
• prompts/scoring_templates.json: the full detection, fuzz, active-fuzz, and intruder templates, including few-shot messages, response instructions, and generation parameters. Template prefixes and user-message structure were checked against real saved calls from each of the 36 layers. The templates show Delphi's few-shot text as the original campaign sent it; for the 543 revised transforms, its <|endoftext|> was shown as ⟪endoftext⟫ (the exact messages are in prompts/cases).
• prompts/scoring_prefixes.json: shared system and few-shot messages, keyed by their SHA-256 identifier.
• prompts/cases/layer-XX-NNNN.jsonl: every recorded scoring request attempt, represented as a shared prefix plus its final user message. Shards are at most 256 MiB, except if an individual record exceeds that limit. All source message sequences were checked for exact reconstruction during export.
• provenance.json: checkpoint revision, per-layer cache hashes, original interpretation fingerprints, method settings, completion times, counts, and revisions, which lists every revised (layer, feature_id).

Read explanations and scores

from datasets import load_dataset

features = load_dataset(
    "georglange/qwen3-4b-molt-interpretations",
    "features",
    split="test",
)
row = features[0]
print(row["layer"], row["feature_id"], row["title"])
print(row["explanation"])
print(row["methods"]["detection"]["metrics"]["accuracy"])

The unique feature key is (layer, feature_id), with zero-based layer and feature IDs. The test split refers to the source activation corpus; the scoring pipeline separately split conversations into explanation and evaluation partitions.

Each feature record contains:

• layer, feature_id, rank, title, explanation, and status.
• activation_count, frequency, maximum, and mean_active: gate activity statistics in that layer's 10M-token sample.
• evidence_counts: actual explanation, positive/negative evaluation and active-fuzz sample counts; short-window count and full-budget flag.
• sampling_exceptions: recorded exceptions to the sampling policy.
• methods: detection, fuzz, active_fuzz, and intruder. Each has status, nullable reason, and metrics.

Metrics include total, valid, parse_failures, accuracy, and accuracy_ci95. Binary methods also have tp, tn, fp, fn, f1, and balanced_accuracy. Intruder has correct and chance. Non-applicable or unavailable metrics are JSON null, not zero. All accuracy-like metrics use a 0–1 scale. accuracy_ci95 is a two-element Wilson confidence interval.

Reconstruct an exact scoring prompt

import json
from datasets import load_dataset
from huggingface_hub import hf_hub_download

repo = "georglange/qwen3-4b-molt-interpretations"
path = hf_hub_download(repo, "prompts/scoring_prefixes.json", repo_type="dataset")
with open(path) as handle:
    prefixes = json.load(handle)

requests = load_dataset(repo, "scoring_prompts", split="test", streaming=True)
request = next(iter(requests))
messages = prefixes[request["prefix_id"]] + request["suffix_messages"]
body = {
    "model": "qwen3.8-27b-nvfp4",
    "messages": messages,
    "max_tokens": request["max_tokens"],
    "temperature": 0,
    "chat_template_kwargs": {"enable_thinking": False},
}

Request records include layer, feature_id, audit_line, task, prefix_id, suffix_messages, and max_tokens. audit_line is the 1-based line number in the original feature's audit. Repeated attempts remain repeated records. Cached calls were not logged again, so request counts need not equal the sum of scored decisions. task is binary or intruder: the original audit used binary for detection, fuzz, and active fuzz, and did not separately label the two fuzz variants. The templates distinguish detection from fuzz; both fuzz variants use identical prompt instructions with different highlights. These request records are not a row-by-row join to aggregate metric counts.

Coverage and method

Feature status	Count
Complete with full evidence budget	80,283
Limited evidence	5,195
Not observed	3,792
Discarded work	10

84,629 transforms have all four valid scores, including some with smaller sample budgets. Discarded intruder work does not remove the other valid scores.

The usual budget is 32 explanation windows, 50 held-out activating windows, 50 held-out zero-gate windows, and a 64-token context cap. Detection and both fuzz methods normally have 100 decisions. Intruder normally has 20 groups, each containing four active contexts and one inactive context. Intruder is a coherence test and does not receive the explanation. Simulation was excluded. Sampling uses seed 42, conversation-level separation, exact token-window deduplication, and peak-ending windows. Sparse features retain smaller sample budgets or explicitly unavailable methods.

The Delphi revision is 4fea06e6e8b68eeaf302474325fca13df95c5d6f. Implementation and adapters: crosslayer-transcoder. The original interpretation fingerprint covers the local run's full settings; it is retained as source provenance, not recalculated from this export.

Interpretation limits

The same model generated explanations and judged them. Binary chance accuracy is 0.5; intruder chance accuracy is 0.2. Negative passages are not semantically matched, and highlight positions can provide cues. Confidence intervals are approximate for paired or reused contexts. Explanations describe scalar gate activation, not the causal effect of the transform on the language model. Transforms outside the special-token revision were explained and scored before the display fix. Their prompts may still contain a rare literal <think>-style string (and always Delphi's few-shot <|endoftext|>), which the scorer tokenized as a control token.

Revisions

special-token-display (October 14, 2026). The Qwen tokenizer turns every added-token string in raw text into its control token, including inside user messages. The original campaign showed <|im_start|>/<|im_end|> evidence tokens as ⟪im_start⟫/⟪im_end⟫, but not <think>, </think>, <tool_call> and similar tokens, nor the <|endoftext|> in Delphi's few-shot examples. Gates that fire on those tokens therefore received explanations that quoted an empty token (for example The token "" appears…) and near-chance scores. Every added token is now shown as plain text, such as ⟪think⟫ and ⟪/think⟫.

543 transforms across all 36 layers were re-explained and rescored with the fix: those with at least 4 of their 8 strongest explanation windows, or at least 25% of their held-out positive windows, peaking on an added token. The evidence, settings and scorer model were unchanged; the explanation, title and all four scores were regenerated, and their scoring requests replace the earlier ones in prompts/cases. Previous values are in dataset revision cae9e4c742f1c1883e28936b8a8959ba03fdf7b4.

Revised transforms, by dominant peak token	n	Mean composite before	After	Near chance before	After
<think> / </think>	57	0.661	0.940	28	1
<|im_start|>	121	0.899	0.913	2	1
<|im_end|>	162	0.952	0.952	1	3
Selected by held-out share only	203	0.775	0.816	46	31
All	543	0.844	0.891	77	36

Composite is the mean balanced accuracy of detection, fuzz and active fuzz; near chance is a composite of at most 0.55. Explanations quoting an empty token fell from 34 to 0. Because every explanation was regenerated, individual transforms also move by ordinary explanation variance: 209 improved by more than 0.02, 228 stayed within 0.02, and 105 declined by more than 0.02, mostly transforms with two unrelated triggers whose new explanation names only one. The <|im_start|>/<|im_end|> groups, already escaped before the fix, show that variance. The intruder score, which never sees the explanation, did not change on average (0.826 to 0.825).

Source texts come from HuggingFaceH4/ultrachat_200k.