b4ph/clap-retrieval-benchmark

CLAP audio-text retrieval, pinned protocol

Retrieval measurements for LAION's fused and unfused CLAP checkpoints on AudioCaps and Clotho, with pinned inputs, per-query results and paired confidence intervals. Under the short-clip protocol, AudioCaps differences are inconclusive. On Clotho, unfused scores higher in both retrieval directions, with the reported 95% intervals excluding zero. Practical equivalence is not established on any endpoint.

The repository also measures how the fused checkpoint's scores depend on processor branch selection and random seeding. It contains evaluation artifacts and code; no audio or model weights are redistributed.

Results

R@1 is the fraction of queries whose top-ranked result is relevant. T2A is text-to-audio; A2T is audio-to-text. Scores below are computed from the recorded per-query ranks. Differences are unfused minus fused, in percentage points (pp).

Endpoint	Fused R@1	Unfused R@1	Difference (pp)	95% CI (pp)
AudioCaps T2A	0.349	0.334	-1.46	[-4.49, +1.57]
AudioCaps A2T	0.338	0.347	+0.94	[-1.88, +3.97]
Clotho T2A	0.132	0.154	+2.24	[+0.86, +3.66]
Clotho A2T	0.154	0.199	+4.50	[+2.01, +7.08]

Source: analysis summary and per-query runs, tags a1_short and a2_short. These are measured runs with explicit short-clip routing, not adjustments to earlier scores.

Intervals use 2,000 paired bootstrap resamples clustered by audio clip, seed 20261028. They describe sampling uncertainty within this run, not variation across training or inference seeds. The summary contains endpoint-wise intervals, not multiplicity-adjusted permutation p-values. None of the 90% intervals lies entirely within the specified equivalence margin of ±3 pp. An interval containing zero therefore does not establish a tie.

Fixtures and protocol

Fixture	Split	Audio clips	Captions / T2A queries	A2T queries
CLAPv2/audiocaps	test	957	957	957
gijs/clotho	test	1,045	5,225	1,045

Counts and fixture fingerprints are recorded in each result file. Clotho uses five separate reference captions per clip. T2A has one relevant audio item per query; Clotho A2T has five relevant captions and counts a hit when any is ranked first. The loader strips surrounding whitespace, removes empty and duplicate captions within each clip, and groups rows by clip identity. See fixture construction.

Both checkpoints use the same fixtures and ranking procedure:

• Audio: 48 kHz mono, soxr resampling, first 10 seconds of each clip.
• Inference: fp32, audio batch 1, text batch 64, processor padding, is_longer=False for the trimmed clips.
• Randomness: Python, NumPy, torch and CUDA seeded with 20261028.
• Ranking: cosine similarity over L2-normalized embeddings; stable sorting resolves ties by document order.

The run files record NVIDIA A40, Python 3.12.3, torch 2.8.0+cu128 and CUDA 12.8. The dependency lock specifies transformers 5.21.0; the full environment record includes inherited container packages.

Checkpoint or fixture	Pinned revision
laion/clap-htsat-fused	365dea6ef167
laion/clap-htsat-unfused	8fa0f1c6d043
CLAPv2/audiocaps	4d18fe131daa
gijs/clotho	93be4f14a2d6

Full commit hashes are in PINS.json. The evaluators require those pins rather than resolving the latest revision.

Processor behavior and checks

In the pinned Hugging Face feature extractor, fusion mode selects one clip with np.random.randint and sets is_longer=True when every clip in a batch fits within the 10-second window. At batch 1, this routes every clip through the long-audio branch. Larger batches also make the selected clip depend on the random state. The current short-clip protocol explicitly sets the flag to False; seeding controls repeatability separately. A fix that removes the assignment, so the flag follows clip length, is proposed in PR #50101. These results set the flag explicitly, so they describe the short-clip branch under either processor behaviour.

On the AudioCaps fixture, the short-clip fused path scores 0.3490 T2A R@1, compared with 0.3417 for the laion_clap control. The calculated difference is 0.73 pp, within the specified 2 pp validation tolerance. Sources: validation summary and package control.

Two seeded processor-native batch-8 runs both score 0.321839 and have identical T2A per-query ranks: run A, run B. This supports repeatability for that tested configuration; the branch choice still affects the score. The processor behavior is documented in transformers issue #50097, with a fix proposed in PR #50101.

An exploratory untrimmed Clotho run scores 0.1233 fused and 0.1435 unfused on T2A R@1. Both are below their trimmed scores. This is one seeded run per checkpoint without intervals: fused, unfused.

Use and reproduction

Read the result summary without downloading audio or running inference:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="b4ph/clap-retrieval-benchmark",
    repo_type="dataset",
    filename="erratum_analysis.json",
    revision="83d9273e683eedfe02162a927e619ceac4f383cd",
)
with open(path, encoding="utf-8") as f:
    summary = json.load(f)
print(len(summary["results"]))  # 4 endpoint records

Each summary record contains the fixture, direction, metric, query and cluster counts, both scores, paired difference, 95% and 90% intervals, and interval-based decision flags. Per-run JSON files contain model and dataset revisions, fixture fingerprints, protocol settings, environment metadata, recall metrics, and per-query first_hit_rank, ap and cluster arrays. Ranks are one-based; scores and intervals are fractions unless displayed as percentage points above.

For inference, use the environment setup and run bash run_erratum.sh from the repository root. The campaign script specifies the short-clip comparison, processor-native repeats and untrimmed arm. The separately recorded batch-8 checks and package control are linked above. The experiment plan records the intended comparisons and decision rules.

Scope and limitations

These results compare two checkpoints on two pinned fixtures. Different audio encodings, reference-caption conventions, windows or preprocessing can change absolute scores, so they should not be compared directly with leaderboard or paper figures. The cause of the low absolute Clotho scores is not fully isolated, and training-data overlap has not been audited here.

The short-clip comparison uses one inference seed. Earlier independent implementation, operator and cross-GPU checks apply to the processor-native campaign. They establish reproducibility under that configuration, but do not independently validate the current short-clip routing. Their evidence remains in REPORT.md, PROVENANCE.json and the independent-operator artifacts.

Protocol version, 2026-10-28: the results above use explicit short-clip routing and seeded inference. Earlier processor-native results, figures and statistical analyses remain available in the original report and analysis. The original plan records the post-collection change to clip-level significance testing; the follow-up plan records the routing update. PUBLICATION.json and SHA256SUMS track the published files.

Repository license: MIT. Upstream audio, captions and checkpoints remain subject to their respective source licenses.