README.md

---
license: cc-by-4.0
pretty_name: LAION Voice Profiles — contrastive DPO pairs (CFG + phase 2)
annotations_creators: [machine-generated]
language_creators: [machine-generated]
language: [en, de]
task_categories: [text-to-speech]
tags: [tts, dpo, preference, rlhf, voice-cloning, speech-emotion, moss, audio-codec, classifier-free-guidance]
size_categories: [100K<n<1M]
configs:
  - config_name: cfg
    data_files: [{split: train, path: cfg/shard-*.parquet}]
  - config_name: p2
    data_files: [{split: train, path: p2/shard-*.parquet}]
---

# LAION Voice Profiles — contrastive DPO pairs (CFG + phase 2)

**Authors: Christoph Schuhmann and LAION.**

**842,935 preference pairs in four families**, built from the same 500 synthetic voice profiles as
[`laion/laion-voice-profiles-sft`](https://huggingface.co/datasets/laion/laion-voice-profiles-sft)
and [`laion/laion-voice-profiles-dpo`](https://huggingface.co/datasets/laion/laion-voice-profiles-dpo).
These are the two pair families that the sister DPO set does **not** contain: they were built later,
for two measured defects of the models trained on it, and they are the complete remainder of the
project's preference material that is publishable under CC-BY-4.0.

| config | family | pairs | contrast | words in the prompt? |
|---|---|--:|---|---|
| `cfg` | `cfg_high` | 237,209 | same voice, matched length, the **top 1 %** of that voice's own range on one measured dimension against the **bottom 1 %**; the instruction asks for the dimension **high**, the high clip is chosen | no — replaced by `...` |
| `cfg` | `cfg_low` | 237,209 | the **same 237,209 pairs with the roles flipped**: the instruction asks for the dimension **low**, the low clip is chosen | no |
| `p2` | `p2_emox` | 309,128 | same voice, matched length, **intense on emotion A against intense on emotion B**; the instruction names A | no |
| `p2` | `p2_len` | 59,389 | the **same clip at its true length against itself cut to 50–75 % or run on to 125–150 %**, with the emotion named | yes — the full timed script |

Both sides of every pair ship as MOSS-Audio-Tokenizer-v2 codes (`codes` = chosen, `rej_codes` =
rejected), together with the reference clip the chosen take was generated against (`ref_codes`).
There is no audio in this repository; the waveforms of every source clip are in
[`laion/laion-voice-profiles-annotated`](https://huggingface.co/datasets/laion/laion-voice-profiles-annotated)
and the join is exact — see *The join*.

---

## Read this first

1. **"CFG" is the project's name for the construction, not a guidance scale.** Nothing here was
   generated with classifier-free guidance and there is no numeric guidance weight anywhere in the
   data. The name comes from the *shape* of the pair: two clips of one voice that differ only in how
   strongly one dimension is expressed, put under one prompt that states the level, and emitted in
   both directions — the same conditional/unconditional contrast that guidance exploits at
   inference, moved into the preference data. `cfg_high` / `cfg_low` mean "the instruction asked for
   the dimension high / low", nothing else.
2. **`uid` is a pair id, not a corpus utterance id.** `cfg<pair_id><h|l>` and
   `p2<pair_id><e|c|x>`. The source clips are named in `cfg/pairs_index.parquet` and
   `p2/pairs_index.parquet` (and, for `cfg`, also in the row's own `caption_tpl`), and *those* ids
   are byte-identical to `uid` in `laion/laion-voice-profiles-annotated`.
3. **Every row is `src = 0`, i.e. one of the 500 published synthetic voice profiles.** No real
   recording, no podcast, no broadcast material contributes a clip, a transcript or a word
   timeline to this set. Checked on every row and every source-clip reference, as an allowlist.
4. **The words are deliberately absent from three of the four families.** A DPO pair shares one
   prompt; when chosen and rejected are different recordings with different words, the words would
   make the preference decidable from the transcript. `cfg_free_text = true` tells the prompt
   renderer to print `...` for every speech chunk while keeping every pause, duration, burst and
   direction tag. The `text` column is still shipped — it is the chosen clip's text, for lookup —
   but it must not go into the prompt of those rows.
5. **These rows are the training-corpus format, not the sister set's 86-column format.** They
   are the exact files the two DPO runs below read, shipped unchanged (31 columns, `zstd`, 128
   shards by `uid` hash). The per-clip annotation stack (40 emotion heads, 57 VoiceNet dimensions,
   captions, speaker embedding) is not repeated here; it is one join away.

---

## Files

| path | count | size | what |
|---|--:|--:|---|
| `cfg/shard-000.parquet` … `shard-127.parquet` | 128 | 3,668,193,909 B | `cfg_high` + `cfg_low`, 474,418 rows, 31 columns |
| `p2/shard-000.parquet` … `shard-127.parquet` | 128 | 2,822,830,198 B | `p2_emox` + `p2_len`, 368,517 rows, same 31 columns |
| `cfg/pairs_index.parquet` | 1 | 8,533,671 B | 237,209 base pairs: `pair_id`, `hi_uid`, `lo_uid`, measured values, percentiles, frames |
| `p2/pairs_index.parquet` | 1 | 9,697,805 B | 368,517 rows: `pair_id`, `ch_uid`, `rj_uid`, percentiles, frames, `mode`, `cut_frames` |
| `code/build_cfg.py`, `code/build_pairs2.py` | 2 | — | the two builders, verbatim, with their design notes in the docstrings |
| `code/mix_cfg.py`, `code/mix_pairs2.py` | 2 | — | how the families were concatenated onto the earlier corpora for training |
| `verify.json` | 1 | — | the verification run reproduced at the end of this card |

Rows are bucketed by `blake2b(uid) % 128`, so every shard is a uniform random sample of its
config and the two families of a config are interleaved. Shard `b` of `cfg/` and shard `b` of
`p2/` are unrelated.

```python
from datasets import load_dataset
cfg = load_dataset("laion/laion-voice-profiles-dpo-cfg", "cfg", split="train")
p2  = load_dataset("laion/laion-voice-profiles-dpo-cfg", "p2",  split="train")
```

---

## What the pairs are, and why

Everything the project's models had been trained on *described* an emotion; nothing ever made two
clips compete on **how strongly** one dimension is expressed. Asked for the 0.90–0.98 percentile of
a named emotion, the round-3 model landed at 0.34. The two builds here are the two attempts to put
intensity itself into the preference signal. Quoted numbers below are from the project protocol
(§16, §I.13 of the technical report) and were re-derived from the shipped files where possible.

### `cfg` — direction contrast on one measured dimension, both ways

For one voice and one dimension, the clip pool is that voice's own training clips (1,199,355
clips over 500 voices in the pool: `src = 0`, not held out, with codes), ranked by the value the
annotation model **measured** on the finished audio — never by what the generating prompt asked
for and never by the clip's intended condition.

* **dimensions**: all **40** emotion heads and **15** VoiceNet ordinal ladders that describe *how*
  something is spoken — `AROU BRGT CLRT DFLU FULL RANG RESP ROUG STNC TEMP TENS VALN VOLT VULN
  WARM`. Identity, recording-quality and content heads (`AGEV GEND RCQL BKGN ESTH EXPL`) are
  excluded: they are not things a performer can be asked to do.
* **the two clips**: top 1 % against bottom 1 % of *that voice's* distribution (floor of 24 clips
  per tail when 1 % is smaller). For emotion heads a separation of at least 0.25 corpus IQR is
  required; realised median corrected percentile of the high clip is **0.998**, of the low clip
  **0.000**, and the smallest high-minus-low gap in the set is 0.73. For VoiceNet ladders the two
  clips sit in different ladder buckets (high side mostly buckets 4–6, low side 0–1).
* **same voice** (`voice_key`), so speaker identity cannot decide the preference; **frames within
  10 %** (`|f_hi − f_lo| ≤ 0.10 · min`, true on 237,209 / 237,209), so length cannot decide it;
  **no words**, so the transcript cannot decide it. Language is **not** matched — 54.6 % of pairs
  are same-language; the row's `lang` is the chosen clip's.
* **nine base pairs per (voice, dimension) cell**, 237,209 base pairs in all (173,116 on emotion
  heads, 64,093 on VoiceNet ladders), 408–495 per voice.
* **both roles.** Every base pair is emitted twice. `cfg_high` (`uid` ends in `h`): instruction
  asks for the dimension high, `codes` is the high clip, `rej_codes` the low one. `cfg_low` (`l`):
  instruction asks for it low, roles swap. Each clip is therefore chosen in one row and rejected in
  another; a policy that simply prefers louder or denser audio scores 50 % on this set, and the only
  way to win is to condition on the instruction. The two families are balanced to the row.

**How the level is stated.** For an emotion head the row carries the corpus's own caption form,
`A voice <band phrase> expressing <emotion>, <prose>; reads as <emotion>`, and `emo_strength` is set
so the prompt renderer picks the matching band:

| role | `emo_strength` | band | band phrases | prose tail |
|---|---|---|---|---|
| high | 0.95 or 0.99, alternating by `uid` hash | `intense` / `extreme` | *intensely, powerfully, viscerally* / *to the extreme, utterly, overwhelmingly* | the corpus's expressed-side prose for that emotion |
| low | the low clip's own percentile, capped at 0.60 (0.00 on 162,160 of 173,116 rows) | `faint` | *mildly, slightly, faintly, a little* | one of three generic "barely there" tails |

The high role alternates between `intense` and `extreme` on purpose: both words are true of a clip
at percentile 0.998, and a two-point contrast cannot calibrate *between* them. For a VoiceNet
ladder there is no emotion to name; `caption_general` is `A voice; <clause> is <ladder tag>` and
`cfg_neutral` carries the direction the renderer puts into the script — `"<tag>, and hold it there;
otherwise exactly as this voice normally speaks"` — with `emo_strength = null` (128,186 rows).

### `p2` — the two phase-2 families

Built after the `cfg` run, for two defects it measured. The base pool and the matching rules are
the same as above; the selection constants are `HI_PCT = 0.90`, `LO_PCT = 0.50`, `FRAME_TOL =
0.10`, 8 emotion-contrastive and 6 length pairs per cell, seed 4711.

**`p2_emox` — intense against intense, for selectivity.** Per-emotion adapters trained on
intense-versus-mild contrasts raised their own emotion by +0.047 when asked for it — and by +0.033
when not. "A ratio of 1.4 : 1 is not control, it is tinting" (the builder's docstring). Every
earlier contrast pits an intense clip against a mild one, so a model can win by being generically
expressive. Here the chosen clip sits at percentile ≥ 0.90 on the head being asked for *and* has
that head as its own top emotion; the partner is ≤ 0.50 on that head and ≥ 0.90 on *its* own top
head. Both sides really are intense, on different emotions, same voice, frames within 10 % (true
on 309,128 / 309,128), words removed. Each base pair is emitted twice, the mirror naming the
partner's own top emotion with the roles swapped (154,564 base pairs, 19,884 (voice, emotion)
cells, all 40 heads). Realised: chosen-side percentile median 0.986 (min 0.900), rejected-side
median 0.215, median gap 0.76.

One honest detail about the mirror: the ≤ 0.50 bound was enforced on the *original* direction
only. In the mirror row the rejected clip is the original's chosen clip, whose percentile on the
partner's emotion was never constrained. Across all 309,128 rows the rejected side is ≤ 0.50 on the
named head on 247,324 (80.0 %), above 0.50 on 61,804, and ≥ 0.90 on 14,281 (4.6 %) — those last
are pairs where one clip is intense on both emotions — and on 3,372 rows (1.1 %) the rejected side
is actually *higher* on the named head than the chosen one (worst gap −0.095). `p2/pairs_index.parquet` carries `ch_pct`
and `rj_pct` per row, so you can filter on the gap you want.

**`p2_len` — the right length, conditioned on the emotion.** The sister set's `truncation` and
`continuation` families know nothing about emotion and teach "the right length is better" in the
abstract. Here the speaker, the text and the emotion stay fixed and only the length moves: chosen
is a clip whose top emotion is at percentile ≥ 0.90, rejected is the **same clip** cut to 50–75 %
of its frames (`mode = cut`, 29,564 rows; realised keep fraction 0.49–0.75, median 0.62) or run on
to 125–150 % by appending the opening frames of another clip of the same voice (`mode = ext`,
29,825 rows; realised 1.23–1.50, median 1.37; the donor is `rj_uid`, same language on 54 %).
Because the words are shared, the prompt carries the full timed script — the timing tags are
exactly what the rejected side violates. Clips shorter than 40 frames (3.2 s) were not used.

### What they trained, and what happened

| adapter | corpus | pairs seen | reward | WER | emotion pct | quality | burst | burst hit rate |
|---|---|--:|--:|--:|--:|--:|--:|--:|
| SFT-3 base, no adapter | — | — | 0.4584 | 0.0987 | 0.3494 | 0.9127 | 0.3564 | 0.666 |
| previous best DPO (sister families only) | 1,853,486 pairs | step 3216 | 0.4687 | 0.1094 | 0.3401 | 0.9211 | 0.3929 | 0.709 |
| [`laion/moss-va-sft3-dpo-lora`](https://huggingface.co/laion/moss-va-sft3-dpo-lora) | sister families + **`cfg`** (2,327,904; 20.4 % `cfg`) | step 4912 | 0.4708 | **0.0950** | 0.3373 | **0.9235** | **0.4271** | **0.772** |
| [`laion/moss-va-sft3-dpo-lora-p2`](https://huggingface.co/laion/moss-va-sft3-dpo-lora-p2) | the above + **`p2`** (2,696,421; 13.7 % `p2`) | step 5022 | **0.4757** | 0.0977 | **0.3541** | 0.9208 | 0.4180 | 0.762 |

Both runs: LoRA rank 64 on
[`laion/moss-tts-local-transformer-4.55b-voice-acting-v2-sft3`](https://huggingface.co/laion/moss-tts-local-transformer-4.55b-voice-acting-v2-sft3),
β 30, lr 1e-6, 256 pairs per step, 8 nodes, stopped by a 6-hour wall clock at roughly half an
epoch; 80-prompt standard evaluation, 320 clips per row.

Read plainly: the `cfg` families made the best *general* checkpoint the project had measured —
word error rate better than the supervised base, the highest quality and burst realisation — and
did **not** move the thing they were built for; emotion percentile fell. Preference accuracy on the
`cfg` rows went 0.56 → 0.98, so the model learned to *recognise* the level without gaining the
ability to *reach* it — the high side of a pair is only as intense as the corpus gets. The `p2`
families then produced the first preference-tuned model above the supervised baseline on emotion
percentile (0.3541 against 0.3494, requested band 0.90–0.98), with emotion-contrastive preference
accuracy 0.951 post-warmup, while `p2_len` saturated at accuracy 1.000 in the second half of the
run — that family is solved and could be dropped from a future mix. The remaining half epoch of
each run was never trained.

---

## The prompt

The rows carry the *ingredients* of the prompt, not the rendered string, because the renderer is
stochastic (direction placement, timed/untimed script, burst dropout, reference-versus-name slot,
all drawn per step). The training code rendered every row into the base model's `<user_inst>`
template with `prompt_lib2.render_prompt` (format hash `073aeb09dc923376`); the renderer itself
is part of the training stack and is not shipped here. What it does with the columns of this set:

* `caption_general` → the `GENERAL:` line (the level statement for emotion rows);
* `text` + `words_json` + `burst_*` + `dur_s` → the timed `SCRIPT:` — segment durations, pauses,
  detected bursts with their lengths, and the delivery direction drawn from `emo_strength`'s band
  (or `cfg_neutral` for VoiceNet rows);
* `cfg_free_text = true` → every speech chunk of that script becomes `...`;
* `frames` → the `- Tokens:` budget; `lang` → `- Language:`; `ref_codes` → the `<|audio|>`
  reference slot, or `Speaker: <speaker_name>` in name mode.

Two rows of this set, rendered by that code (reference-audio mode):

`cfg_high`, emotion head *Intoxication*, uid `cfgd0ac06bc0eea0a5f4dcch`:

```
<user_inst>
- Reference(s):
<|audio|>
- Instruction:
GENERAL: A voice utterly expressing intoxication, a woozy, slurring voice, intoxicated and untethered, in an altered haze, impossible to hide; reads as intoxication altered states of consciousness
SCRIPT:
(not holding it together at all, breath shallow and fast, the fear plain in the voice, unguarded; completely intoxicated) [4.3 seconds duration] ... [0.5 seconds pause] (keep it overwhelmingly intoxicated) [2.7 seconds duration] ...
- Tokens:
94
- Quality:
None
- Sound Event:
None
- Ambient Sound:
None
- Language:
English
- Text:
(not holding it together at all, breath shallow and fast, the fear plain in the voice, unguarded; completely intoxicated) [4.3 seconds duration] ... [0.5 seconds pause] (keep it overwhelmingly intoxicated) [2.7 seconds duration] ...
</user_inst>
```

`cfg_low`, VoiceNet ladder *WARM* (timbre), uid `cfga19aa3876e64686f91c1l` — no `GENERAL:` line,
the level sits in the script as `cfg_neutral`:

```
- Instruction:
SCRIPT:
[5.5 seconds pause] (slightly cool, and hold it there; otherwise exactly as this voice normally speaks) [1.4 seconds duration] ... [0.8 seconds pause] [3.2 seconds duration] ... [0.3 seconds pause]
- Tokens:
140
```

A `p2_len` row renders the same way with the words present in the script. The `[n seconds
duration]` and `[n seconds pause]` tags come from `words_json`; the `(burst, n seconds)` tags from
`burst_starts` / `burst_ends` / `burst_labels`; the parenthesised directions are drawn from the
band and are not stored in the row.

---

## Columns — the 31 columns of `cfg/shard-*` and `p2/shard-*`

| column | type | meaning |
|---|---|---|
| `uid` | `string` | pair-row id. `cfg` + 20-hex `pair_id` + `h`/`l` (role), or `p2` + 9-digit `pair_id` + `e`/`c`/`x` (`emox` / `cut` / `ext`). Unique across the whole set. |
| `family` | `string` | `cfg_high`, `cfg_low`, `p2_emox`, `p2_len` |
| `src` | `int8` | source group of the training corpus; **0 on every row** = synthetic voice profile |
| `voice_key` | `string` | voice-profile id, one of the 500 (`anime_000`, `emolia_c0123`, `mediathek_…`, `refvoice_…`, `k<n>_age<n>_bg<n>`); both clips of the pair share it |
| `lang` | `string` | `en` / `de` — of the **chosen** clip; the rejected clip of a `cfg` / `p2_emox` pair may be in the other language |
| `text` | `string` | the chosen clip's generation text, verbatim (may contain the generator's own parenthesised burst directions). **Not** in the prompt when `cfg_free_text` is true. |
| `text_bursts` | `string` | `text` with the *detected* vocal bursts inserted inline as `(label)`; equals `text` when nothing was detected (73 % of `cfg` rows, 75 % of `p2`) |
| `caption_general` | `string` | the level statement. Emotion rows: `A voice <band> expressing <emotion>, <prose>; reads as <emotion>`. VoiceNet rows: `A voice; <clause> is <tag>`. |
| `caption_script` | `string` | always empty in this set |
| `caption_tpl_text` | `string` | always empty in this set |
| `caption_tpl` | `string` | provenance key. `cfg`: `cfg\|<emo\|vn>\|<dimension>\|<hi_uid>\|<lo_uid>`. `p2`: `p2\|<emox\|len>\|<emox\|cut\|ext>\|<dimension>` |
| `speaker_name` | `string` | the voice's assigned given name (the same mapping as `names.csv` in the sister set) |
| `spoken_dur_s` | `float32` | end of the last aligned word of the chosen clip, seconds |
| `dur_s` | `float32` | duration of the chosen clip, seconds (generator's pre-encode measurement; `frames == round(dur_s · 12.5)`) |
| `words_json` | `string` | JSON list `[{"w","s","e"}, …]` — MMS_FA forced alignment of the chosen clip's text, seconds |
| `burst_starts` / `burst_ends` | `list<float32>` | detected vocal-burst spans in the chosen clip, seconds |
| `burst_labels` | `list<string>` | their class labels (`Contented Sigh`, `Breathy Giggle`, …) |
| `in_extreme` | `bool` | legacy flag of the older corpus; **always false** here (`emo_strength` carries the band instead) |
| `frames` | `int32` | MOSS frames of the chosen clip |
| `codes` | `binary` | **chosen** MOSS codes, little-endian `uint16`, `frames × 12` values |
| `rej_frames` | `int32` | MOSS frames of the rejected side |
| `rej_codes` | `binary` | **rejected** MOSS codes, same layout. `cfg` / `p2_emox`: the other clip; `cut`: a frame-aligned prefix of `codes`; `ext`: `codes` followed by `cut_frames` frames of the donor |
| `ref_frames` | `int32` | MOSS frames of the reference clip |
| `ref_codes` | `binary` | the reference clip the chosen take was paired with in the training corpus — another take of the same voice; its uid is not carried in these rows |
| `has_ref` | `bool` | always true here |
| `is_val` | `bool` | `blake2b(uid) % 1000 == 0` — the training runs' held-out rows (490 in `cfg`, 373 in `p2`); shipped, not excluded |
| `cfg_free_text` | `bool` | render the script without words. True on every `cfg` row and every `p2_emox` row, false on `p2_len` |
| `cfg_neutral` | `string` | VoiceNet rows only: the direction stating the ladder tag (`"very tense, and hold it there; otherwise exactly as this voice normally speaks"`); empty elsewhere |
| `emo_strength` | `float32`, nullable | the emotion band the prompt states, as a percentile: 0.95 / 0.99 for the high or named side, ≤ 0.06 for `cfg_low` emotion rows, **null** on VoiceNet rows |

### `cfg/pairs_index.parquet` — 237,209 rows, one per base pair

| column | type | meaning |
|---|---|---|
| `pair_id` | `string` | 20 hex characters; the row uids are `cfg{pair_id}h` and `cfg{pair_id}l` |
| `voice_key` | `string` | the voice |
| `dim` | `string` | emotion head name or VoiceNet code |
| `kind` | `string` | `emo` (173,116) or `vn` (64,093) |
| `hi_uid` / `lo_uid` | `string` | the high and the low clip — **uids of `laion/laion-voice-profiles-annotated`** |
| `hi_val` / `lo_val` | `float32` | the measured value of `dim` on each clip (emotion intensity, or VoiceNet regression) |
| `hi_frames` / `lo_frames` | `int32` | their MOSS frames |
| `hi_pct` / `lo_pct` | `float32` | `emo`: tie-aware corpus ECDF percentile of the value; `vn`: the ladder bucket (0–6) as a float |

### `p2/pairs_index.parquet` — 368,517 rows, one per shipped row

| column | type | meaning |
|---|---|---|
| `pair_id` | `string` | 9 digits; the row uid is `p2{pair_id}{mode[0]}` |
| `family` | `string` | `emox` or `len` (the row's `family` is `p2_` + this) |
| `voice_key`, `dim` | `string` | the voice; the emotion head the instruction names |
| `ch_uid` | `string` | the chosen clip — uid in `laion/laion-voice-profiles-annotated` |
| `rj_uid` | `string` | `emox`: the rejected clip; `cut`: equal to `ch_uid`; `ext`: the donor whose opening frames are appended |
| `ch_pct` / `rj_pct` | `float32` | percentile of `dim` on the chosen / rejected clip (`len`: both the chosen clip's own top-head percentile) |
| `ch_frames` / `rj_frames` | `int32` | `emox`: frames of the two clips; `cut`: `rj_frames` is the kept prefix length; `ext`: `rj_frames` is the number of donor frames appended |
| `mode` | `string` | `emox` / `cut` / `ext` |
| `cut_frames` | `int32` | `cut`: frames kept; `ext`: frames appended; `emox`: 0 |

---

## The join

`hi_uid`, `lo_uid`, `ch_uid`, `rj_uid` **are** the `uid` column of
[`laion/laion-voice-profiles-annotated`](https://huggingface.co/datasets/laion/laion-voice-profiles-annotated)
(`index/origin=original/`, and the tar member stem under `data/origin=original/`), of the form
`<voice>__<block>__<cell>__<lang>.cNNN` — e.g. `anime_000__E__Affection__D__de.c037`. Verified
byte-for-byte on the released index (see below). From there you have the MP3, the 205-column
annotation, the speaker embedding, and — via the same uid — the rows of
`laion/laion-voice-profiles-sft` and `-dpo`.

```python
import numpy as np, pyarrow.parquet as pq, pyarrow.dataset as ds

def codes(b, frames, n_vq=12):
    return np.frombuffer(b, dtype="<u2").reshape(frames, n_vq)

row = pq.read_table("cfg/shard-000.parquet").slice(0, 1).to_pylist()[0]
chosen, rejected = codes(row["codes"], row["frames"]), codes(row["rej_codes"], row["rej_frames"])

# which clips are these?  (cfg: also readable from caption_tpl.split("|")[3:5])
idx = pq.read_table("cfg/pairs_index.parquet").to_pandas().set_index("pair_id")
p   = idx.loc[row["uid"][3:-1]]
chosen_uid, rejected_uid = (p.hi_uid, p.lo_uid) if row["uid"].endswith("h") else (p.lo_uid, p.hi_uid)

# the same for p2
row = pq.read_table("p2/shard-000.parquet").slice(0, 1).to_pylist()[0]
p   = pq.read_table("p2/pairs_index.parquet").to_pandas().set_index("pair_id").loc[row["uid"][2:-1]]
chosen_uid, rejected_uid, mode = p.ch_uid, p.rj_uid, p.mode     # rj_uid == ch_uid when mode == "cut"

# audio + full annotation, from the annotated corpus (local snapshot)
ann = ds.dataset("laion-voice-profiles-annotated/index", partitioning="hive", format="parquet")
a   = ann.to_table(filter=ds.field("uid") == chosen_uid).to_pylist()[0]
# tar: data/origin=original/{a['shard']}.tar, members {chosen_uid}.mp3 / .json / .moss.npy / .vclap.npy
```

`uid` in the annotated corpus **contains dots** — do not split on the first one. Its
`audio_key` is not unique across runs; join on `uid`.

---

## MOSS codes

`codes`, `rej_codes`, `ref_codes` are raw little-endian `uint16`, `(frames, 12)` after reshape:
**12 codebooks × 1024 entries at 12.5 fps, one frame = 12 tokens = 80 ms.** There is no 32-token
block and no interleaving; the 32 associated with this stack is the codec's `num_quantizers`, of
which the TTS model consumes the first 12. Every `cut` boundary is a whole frame. The `p2` builder drops
rejected sides longer than 620 frames (49.6 s); the training corpus caps chosen takes the same way.

**These codes are from the corrected tokenisation.** The voice-profile corpus once carried codes
computed on a half-speed decode (a stereo-to-mono bug that doubled every duration and frame count
in the first annotation tree). The frame counts here were compared, clip by clip, with
`moss_frames` of the released annotated index for 1,322 clip references of voice `anime_000`
(its first three index parts): ratio **1.00 on every one**, and `dur_s` likewise. Nothing in this set comes from the broken tree.

---

## The material this is built from

500 synthetic **voice profiles**, each one reference speaker driven through a fixed matrix of 842
named acting conditions in English and German, keeping every candidate take. Generator
[`laion/moss-tts-local-transformer-4.55b-voice-acting-v2`](https://huggingface.co/laion/moss-tts-local-transformer-4.55b-voice-acting-v2),
codec [`OpenMOSS-Team/MOSS-Audio-Tokenizer-v2`](https://huggingface.co/OpenMOSS-Team/MOSS-Audio-Tokenizer-v2),
run `vprof_base`. The full corpus with audio and the complete annotation stack is
`laion/laion-voice-profiles-annotated`; the per-voice LoRAs and reference clips are
[`laion/moss-voice-profile-loras-500`](https://huggingface.co/laion/moss-voice-profile-loras-500).
The clip pool for both builds is the voice-profile part of the project's second training corpus
(`src = 0`, 1,199,355 clips) — the same top-3-per-cell takes that make up
`laion/laion-voice-profiles-sft` (1,200,531 rows), minus the corpus's 1-in-1000 held-out rows, joined
with the corrected re-annotation. The source clips span every block of the matrix — for
the `cfg` high side, 115,656 come from emotion cells, 88,904 from VoiceNet cells, 22,562 from edge
cases, 6,532 from character voices, 2,957 from isolated bursts, 380 sports and 218 explicit cells —
because the ranking is by the measured value, not by the cell.

The measured values that rank the clips are the 40 Empathic-Insight emotion intensities and the 57
VoiceNet regressions of the corrected re-annotation (`laion/voiceclap-commercial` →
`laion/voicenet-dimension-predictors-commercial`), with the percentiles taken from the project's
tie-aware ECDF over 132.8 M rows pooled across its corpora. The same caveat as on the sister cards
applies: these are model outputs, not human ratings, and the VoiceNet heads were trained on
Gemini perceptual estimates (mean r 0.79).

---

## Limitations, honestly

* **No human has listened to any pair in a controlled study.** "High" and "low" are model
  measurements on model output.
* **The high side is only as intense as the corpus gets.** These pairs teach the *direction* of a
  dimension within a voice's own range; they cannot teach a level the generator never produced.
  That is exactly the result the `cfg` run showed.
* **Language is not matched within a `cfg` / `p2_emox` pair** (54.6 % / 59.1 % same-language).
  The prompt states the chosen clip's language and the words are absent, so the mismatch is not
  visible in the prompt — but the rejected side may be a German clip against an English prompt.
* **The `p2_emox` mirror rows are looser than the original direction** (20.0 % of rows with the
  rejected side above 0.50 on the named head; 4.6 % at or above 0.90). Filter on `rj_pct` if you
  want the strict version.
* **`p2_len` negatives are constructed, not observed**, and the `ext` splice has no crossfade.
  Its preference accuracy saturated at 1.000 during training; it carries little signal for a model
  that already handles timing tags.
* **The rendered prompts are not shipped**, only their ingredients; reproducing them exactly needs
  the project's renderer. The example strings above are what the trainer saw.
* **Speaker identity is weak against the nominal reference, everywhere**, as inherited from the
  generation run (per-voice mean `spk_sim` 0.18–0.69 in the sister set). Same voice within a pair
  is guaranteed by `voice_key`, not by a similarity threshold.

---

## What is, and is not, in the sister sets

| repository | families | relation to this set |
|---|---|---|
| [`laion/laion-voice-profiles-dpo`](https://huggingface.co/datasets/laion/laion-voice-profiles-dpo) | `emotion`, `truncation`, `continuation` — 3,451,531 pairs | the `vp_*` families of the training corpora; **none of the four families here** |
| [`laion/tts-realspeech-dpo-en-de`](https://huggingface.co/datasets/laion/tts-realspeech-dpo-en-de) | `too_short`, `too_long` on real speech — 3,959,192 pairs | the `rs_*` families; real recordings, disjoint by construction (`src ≠ 0`) |
| **this set** | `cfg_high`, `cfg_low`, `p2_emox`, `p2_len` — 842,935 pairs | the remainder |

The training corpora were nested unions: `dpo_corpus2` (1,853,486, sister families) ⊂
`dpo_corpus_cfg` (+ `cfg`, 2,327,904) ⊂ `dpo_corpus_p2` (+ `p2`, 2,696,421). `code/mix_cfg.py` and
`code/mix_pairs2.py` are the two concatenations. Overlap was checked, not assumed:
every `uid`,
`chosen_uid`, `rejected_uid`, `ref_uid` and `donor_uid` of all 1,922 parquet files of both sister sets
(7,410,723 rows) was read, and **none of the 237,209 `cfg` base pairs and none of the 154,564
`p2_emox` base pairs is a `(chosen, rejected)` tuple of the sister `emotion` family (1,064,594
tuples), in either order.** What *is* shared is the pool: all 332,576 / 342,284 source clips appear
in the sister set as a chosen take, because both are built from the same top-3 selection, and every
`p2_len` chosen clip also has a `truncation` and a `continuation` negative there — cut at a sentence
boundary in the 30–85 % window rather than at 50–75 % of the frames, so the pairs differ. No row of
this set is a copy of a row there.

---

## Verification run on the shipped files

| check | result |
|---|---|
| rows shipped | `cfg` **474,418** (`cfg_high` 237,209 · `cfg_low` 237,209) · `p2` **368,517** (`p2_emox` 309,128 · `p2_len` 59,389) — counted with `pyarrow` over all 256 shards |
| languages (of the chosen clip) | `cfg` en 227,068 / de 247,350 · `p2` en 179,649 / de 188,868 |
| `uid` unique across each config | 474,418 / 474,418 · 368,517 / 368,517 |
| `src == 0` | 474,418 / 474,418 · 368,517 / 368,517 |
| `voice_key` is one of the 500 published profiles | 474,418 / 474,418 · 368,517 / 368,517 (500 voices in each config) |
| every source-clip reference is `<voice>__<block>__<cell>__<lang>.cNNN` with `voice` in the 500 | 1,423,254 / 1,423,254 (`cfg`, index + `caption_tpl`) · 737,034 / 737,034 (`p2`); **0** matches to any real-speech or broadcast uid pattern |
| every row has an entry in its `pairs_index.parquet` | 474,418 / 474,418 · 368,517 / 368,517 |
| sample join, 200 rows per family: role, `voice_key`, `frames`, `rej_frames` agree with the index | 200 / 200 on `cfg_high`, `cfg_low`, `p2_emox`; 82 / 82 on `p2_len` |
| code buffer length `== frames × 12 × 2` for `codes`, `rej_codes`, `ref_codes` | 682 / 682 sampled rows |
| all codes in 0..1023 | 682 / 682 |
| `abs(frames − round(dur_s × 12.5)) ≤ 1` | 682 / 682 |
| `cut`: `rej_codes` is a frame-aligned prefix of `codes` | 41 / 41 sampled |
| `ext`: `rej_codes` begins with all of `codes` | 41 / 41 sampled |
| frame tolerance `|f_hi − f_lo| ≤ 0.10 · min` | 237,209 / 237,209 `cfg` base pairs · 309,128 / 309,128 `p2_emox` rows |
| frames and `dur_s` against the released annotated index (`moss_frames`, `dur_s`) | 1,322 / 1,322 clip references at ratio 1.00 — the corrected tokenisation |
| `has_ref` | true on every row |
| `in_extreme`, `caption_script`, `caption_tpl_text` | always false / empty / empty, all rows |
| overlap with `laion/laion-voice-profiles-dpo` and `laion/tts-realspeech-dpo-en-de` | 7,410,723 sister rows read; **0** identical pairs; 0 / 800 sampled row uids present there |
| files on the Hub vs. local, `HfApi.repo_info(files_metadata=True)` | **258 / 258** files, size and LFS sha256 identical, **6,509,255,583 bytes** |

Each check re-derives a property from a quantity other than the one the writer used — frames
against `dur_s`, code length against the declared frame count, the row's frames against the pair
index, the pair index against the released annotated index — rather than re-running the builders.

---

## Licence and attribution

**CC-BY-4.0.** Credit **Christoph Schuhmann** and **LAION**, and the upstream sources listed in
`laion/laion-voice-profiles-annotated`. Every clip referenced here is model output from one of the
500 synthetic profiles; no take is a recording of a person, and no clip, transcript or timeline from
the project's real-speech or broadcast corpora is included.

```bibtex
@misc{schuhmann2026voiceprofilesdpocfg,
  title  = {LAION Voice Profiles -- contrastive DPO pairs (CFG + phase 2)},
  author = {Schuhmann, Christoph and LAION},
  year   = {2026},
  url    = {https://huggingface.co/datasets/laion/laion-voice-profiles-dpo-cfg}
}
```