76 lines
4.4 KiB
Markdown
76 lines
4.4 KiB
Markdown
|
# Music metadata hypothesis: actual Google CPU diagnostic
|
||
|
|
|
||
|
|
This known two-file diagnostic compares real audio evidence with real embedded
|
||
|
|
metadata. It does **not** change production ranking or establish independent
|
||
|
|
music quality, general abstention, or a higher media rating.
|
||
|
|
|
||
|
|
The preserved v1 run encoded eleven default audio windows and two whole-tag
|
||
|
|
passages. Adding the whole passage left music memories at **2/4 first hits**:
|
||
|
|
the violin's English similarity increased, but stayed below the wrong piano.
|
||
|
|
|
||
|
|
The subsequent v2 policy was frozen in source commit `7055004` before inference.
|
||
|
|
It encodes artist, title, album, genre and composer separately, in fixed order
|
||
|
|
with a shared 2048-byte UTF-8 budget and the neutral document title `File metadata`.
|
||
|
|
The file score is the maximum of its actual evidence, with no weights or cutoff.
|
||
|
|
This is a hypothesis motivated by v1, not untouched held-out testing.
|
||
|
|
|
||
|
|
## Recomputed results
|
||
|
|
|
||
|
|
All 65 original queries remain in the report. Three filter queries are explicitly
|
||
|
|
unsupported. Among 62 executed queries, only seven positive queries name one of
|
||
|
|
the two candidates; 50 positives name files outside this deliberately limited
|
||
|
|
pool and cannot receive a quality score. Five unfiltered no-answer queries remain.
|
||
|
|
|
||
|
|
| Two-file diagnostic | Audio only | Whole tags v1 | Individual fields v2 |
|
||
|
|
| --- | ---: | ---: | ---: |
|
||
|
|
| Music memories: correct first file | 2/4 | 2/4 | 4/4 |
|
||
|
|
| Music identity queries: correct first file | 2/3 | 2/3 | 3/3 |
|
||
|
|
| All eligible positives: correct first file | 4/7 | 4/7 | 7/7 |
|
||
|
|
| Eligible positives: recall within first five | 7/7 | 7/7 | 7/7 |
|
||
|
|
| Eligible positive MRR | 0.786 | 0.786 | 1.000 |
|
||
|
|
| No-answer queries returning candidates, failure | 5/5 | 5/5 | 5/5 |
|
||
|
|
|
||
|
|
The first-hit gains are `q049`, `q050` and `q056`, with no scored positive
|
||
|
|
regression. English violin selected the literal composer field
|
||
|
|
`Violinist WTD/LPC Jazz Band` (cosine 0.73714); German violin selected the actual
|
||
|
|
title `Jazz Violin Solo` (0.73508). Their wrong piano audio maxima were 0.68972
|
||
|
|
and 0.70654 respectively. The piano memories retained their original audio hits.
|
||
|
|
|
||
|
|
This improvement also strengthens unrelated nearest neighbors: the purple-elephant
|
||
|
|
violin negatives increased by 0.04994 and 0.04801 cosine, and the German quantum
|
||
|
|
entanglement negative by 0.00749. All negatives still return files. Consequently,
|
||
|
|
the diagnostic provides **no abstention improvement** and no safe rejection
|
||
|
|
threshold. Metadata is useful evidence, not a calibrated confidence score.
|
||
|
|
|
||
|
|
## Inputs, execution and limits
|
||
|
|
|
||
|
|
The actual piano tags identify Kimiko Ishizaka, Goldberg Variation 13, Classical
|
||
|
|
and J.S. Bach. The violin tags identify Violinist WTD, Jazz Violin Solo, Jazz and
|
||
|
|
the LPC Jazz Band. No instrument descriptions were generated from the waveform.
|
||
|
|
Blank/missing whitelisted fields add no vector; encoder, comments, URLs and
|
||
|
|
copyright are excluded. Misleading tags remain unverified literal evidence.
|
||
|
|
Metadata citations have a field and text, never an invented playback timestamp.
|
||
|
|
|
||
|
|
V2 used actual Google EmbeddingGemma 2 Q8 through the same native
|
||
|
|
`b11461-no-logits-v1` worker on CPU, one thread, 768 dimensions and a 2048-token
|
||
|
|
indexing context. Ten field encodes took **750.03 ms**, excluding model startup
|
||
|
|
and query inference. It made **zero audio inference calls**: per-query audio
|
||
|
|
maxima were reused verbatim from the SHA-verified v1 report. Independent read-only
|
||
|
|
recomputation checked query equality and exact retained audio/whole-tag rankings
|
||
|
|
for all 62 executed rows. This tiny shared-VM measurement is not a throughput
|
||
|
|
guarantee or a production ANN/hybrid measurement.
|
||
|
|
|
||
|
|
- Frozen v2 plan SHA-256:
|
||
|
|
`566bcd608ccf898abaceab32b247786e895368c8428b1bfa6e0bbb8396211b05`.
|
||
|
|
- Preserved v1 report SHA-256:
|
||
|
|
`f985ba639a14b8c0288b8fa7812d580f179f31c8cad0c21d67531c8a4fddc5b6`.
|
||
|
|
- Ignored v2 report `.searchi/qa/music-fields-v2-results.json`, 288944 bytes,
|
||
|
|
SHA-256 `c13ceca2150059a21e10ee190893a66f5ac82570675e7725e54524c660a2ff30`.
|
||
|
|
|
||
|
|
Reproduction commands and structural tests are in the
|
||
|
|
[evaluation tool README](../tools/media-eval/README.md). Before deployment,
|
||
|
|
evaluate independently frozen unseen music, missing/misleading metadata, additional
|
||
|
|
instrument/style hard negatives and the complete mixed candidate pool. Production
|
||
|
|
integration must retain real metadata field citations, source-byte/revision guards,
|
||
|
|
strict filters and generation policy invalidation. The broader
|
||
|
|
[media quality gates and experimental rating](MEDIA-QUALITY.md) remain unchanged.
|