utkarshUY/CharacterPlex-Synthetic-Data-v4

CharacterPlex Synthetic Data v4 (research only)

The synthetic part of the CharacterPlex v4 training corpus: character sheets, LLM-written scenario scripts with director NOTES, per-agent training views, rendering and QC records, word alignments, v4 sidecars (event edits, episode weights, held-out checks) and Mimi token shards of the Breeze TTS 2 renders. CharacterPlex is a full-duplex speech-to-speech model (PersonaPlex/Moshi family) that plays a character from a sheet and follows a director's in-world NOTES.

This repository is not the full v4 training corpus. About 66 % of the v4 text-loss gradient comes from third-party human corpora (Seamless Interaction, AMI, ICSI, Expresso, EARS), which are not included here. No audio waveform and no voice reference is included. See What is NOT included.

INVENTORY.json lists every file in this repo (bytes, sha256, source path on cluster storage), and inventory/members.jsonl.gz lists every file inside each tar shard (shard, path, bytes, sha256, source path; for token shards also the sha256 of the original npz and the removed keys). The excluded categories, with sizes and cluster locations, are also in INVENTORY.json → excluded.

Quick look (browse it in the Dataset Viewer above)

Two small, readable tables sit in sample/ so you can see the format without downloading the shards:

Table	Rows	What one row is
episodes_sample (sample/episodes.jsonl)	12	one training conversation (4 each from the three generators: people, scene2, arc), all from the training split and rated ≥ 4/5 "in character" by the QC judge
model_inputs_sample (sample/model_inputs.jsonl)	50	the exact conditioning text the model was trained on for one character window: the v2 CARD, SCENE, ROSTER and an example NOTES line (the 50 golden windows the trainer checks byte for byte)

sample/raw_episode_v3-people-001648-1.json is one complete episode file, unflattened (schema s2s-episode/1), as stored in data/episodes/.

How an episode is made.

1. A character sheet in the Character Studio format (fabric-character-library-v1) is generated: who, history, goal, obstacle, six personality sliders (0–100, the frontend's polarity), memories, triggers, relationships, voice description and a five-beat story arc.
2. An LLM writes a multi-party scene around it. Every turn carries emotion tags ([e:anger i:3 pace:fast]), nonverbal and pause markers ([laugh], [pause], [long_pause]). The scene also includes a director plan: in-world notes and beat changes the character must act on, with the expected behaviour and the turns that should comply.
3. A judge model scores each script (in character, natural speech, emotion fit; no assistant voice, meta leaks or invented facts), and failures are rejected.
4. Breeze TTS 2 renders every turn in a cloned voice; the audio is aligned word by word to a 12.5 Hz timeline and tokenised with Mimi (8 codebooks), giving data/tokens/.
5. At training time each character's view is turned into the pinned memory the model sees: CARD (the v2 card, ≤ 700 tokens, compiled from the sheet), SCENE, ROSTER, and a NOTES slot that is rewritten mid-window when a director note, beat change or memory arrives. The model_inputs_sample table shows real examples.

Column guide, episodes_sample.

• Who and where: character, role, setting, situation, story_arc, other_participants.
• Personality: personality_traits_frontend_0to100. A low value means the slider's left label (quiet, cautious, practical, serious, orderly, private); the v2 card writes them as words.
• Sheet: character_sheet_json is the full sheet. training_card_text is the card stored with the episode at generation time. The model was trained on the v2 render of the same sheet (see model_inputs_sample).
• Direction: director_plan lists each live note or beat, its target and the expected behaviour.
• Dialogue: turns holds speaker, kind (turn or backchannel) and text with tags.
• Run facts: n_turns, duration_s (rendered length), qc_judge and licence.

What the trained model accepts at runtime

Input	How it reaches the model	Live during a call?
Character profile (v2 card from the sheet: identity, history, personality words, goals, relationships, memories, triggers, voice description)	pinned CARD segment, prefilled at session start	set at start. Can be re-pinned in place only if the new card is the same length or shorter (card op); in practice, change it between sessions
Scene and situation	pinned SCENE	at start (re-pinned with the card)
Who else is present	pinned ROSTER (<slot1> …)	at start (re-pinned with the card)
Director notes, story-beat changes, retrieved character memories, the user's facts	the pinned NOTES slot (64 frames), rewritten live in ~20 ms (notes op), as trained with mid-window NOTES switching	yes
Voice	a voice prompt (stock PersonaPlex voice embeddings) chosen at session start	no, start only. There are no affect variants such as calm vs panicked
The user's speech	the user audio stream, continuously (full duplex)	yes
Forced words, word bans, listen-only	say, ban, quiet control ops (runtime tools, not trained inputs)	yes
Numeric emotion or prosody parameters (pitch, energy, intensity)	not supported. Traits become words on the card, and delivery is steered by NOTES text. The model's own text stream can emit reserved event and delivery tokens: ‹laugh›, ‹chuckle›, ‹sigh›, ‹gasp›, ‹cry›, ‹sniff›, ‹cough›, ‹breath›, ‹groan›, ‹scream›, ‹sob›, ‹tearful›, ‹whisper›, ‹pause›, ‹long_pause›. The v4 data supervised laughs, sighs, whispers, tearful delivery and pauses, but had no labels for ‹groan› or ‹sob› and only 91 for ‹scream›	–

Licence and use

Research and non-commercial use only. The data inherits the following terms:

Component	Terms
Audio-derived tokens (rendered with BreezeBlue/Breeze-TTS-2)	BreezeBlue Research and Non-Commercial Licence v1.1. A model trained on these outputs is a Derivative Model under that licence (§1.3/§5(b)/§7.2); if distributed it must carry its NOTICE and "Derived from Breeze TTS 2".
Voices the renders were cloned from (references not included)	Seamless Interaction (naturalistic) CC BY-NC 4.0, EARS CC BY-NC 4.0, Expresso CC BY-NC 4.0, L2-ARCTIC CC BY-NC 4.0, VCTK CC BY 4.0
Spliced non-verbal clips inside some renders (laugh, sigh, gasp, cough, sniff, cry)	EARS CC BY-NC 4.0, VocalSound CC BY-SA 4.0
Scripts, sheets, NOTES (text)	written by Qwen3.5-122B-A10B-GPTQ-Int4 (Apache-2.0); judged by Gemma 4 31B-it (Gemma terms)
Word timings	Qwen/Qwen3-ForcedAligner-0.6B (Apache-2.0)
Mimi codec	PersonaPlex tokenizer tokenizer-e351c8d8-checkpoint125 (Kyutai Mimi; see PersonaPlex terms)
Pilot-era episodes (episode_gen 1.1-2.3, 1,288 episodes, flag_fish_or_maya=1)	provenance also lists Fish Audio Research Licence (NC), Maya1 (Apache-2.0) and, for 181 2.3-api episodes, Fish Audio API Terms (free tier) from earlier voice design / renders; final audio is Breeze TTS 2

Per-episode licence inheritance and voice-reference corpora are in metadata/episodes_flags.tsv. The model licence table of the project is repo/docs/MODELS_AND_LICENSES.md. Licence details for the voices: the renders are voice clones of real speakers from the corpora above. The token shards can be decoded with Mimi back to audio in those voices. Use only for research, and do not try to identify the speakers.

Contents and counts

All 13,771 episodes of the full-run / v4 episode manifest (metadata/fullrun/episodes/manifest.jsonl, sha256 pinned in metadata/v4/MANIFEST.json). Every episode is synthetic (source: s2s-episodes).

Split	Episodes	Conversation hours
train	10,405	776.2
heldout_voice	2,263	171.7
heldout_sheet	1,046	80.2
heldout	57	3.8
total	13,771	1,031.9

Train target-track hours: 1,036.8 (107,409 unique 3,000-frame windows). Speakers per episode: 2 (12,080), 3 (1,506), 4 (185).

Episode types (id prefix):

Generator	Type	Episodes
gen_v3 3.0	v3-arc 5 arc beats via NOTES	2,876
	v3-trigger sheet trigger, reaction, settle	2,018
	v3-memory memories retrieved into NOTES, distractors	1,980
	v3-script must-hit scene beats/lines	1,925
	v3-scene2 two sheet characters	1,653
	v3-people	1,481
	v3-open free conversation, 0-2 nudges	550
episode_gen 1.1-2.3 (pilot / v2.x)	pl, p2a, p2b, p22a, p23a	1,288

v3 generation pool: 6,701 character sheets (s3-*), 12,506 accepted scripts, 7,373 rejected scripts (judge / repair failures, kept as QC records).

v4 corrections (counts in metadata/v4/MANIFEST.json → corrections): special pause tokens; Breeze event labels energy-verified, moved or dropped (e.g. sigh kept 14,332 / dropped 6,349; laugh kept 5,660 / dropped 4,834); 277 tearful and 14,413 whispered turns; behaviour-coverage episode weights (qa 7,002, recall 2,345, arc 3,753, yield 781; mean weight 1.33); held-out check clean (0 overlap on ids, npz, views, sheets, voices, names; the known exception: 50 held-out character voices are used as the scripted user voice).

Layout

Path	What
README.md, INVENTORY.json, inventory/members.jsonl.gz	this card; full file inventory
metadata/v4/	MANIFEST.json (frozen v4 manifest, original), episode_weights.tsv, heldout_check_v4.json, event_edits_v4.synthetic.json, acc_v4.synthetic.json, examples_v4.synthetic.json, code/ (GPU event-energy scripts), work/ (Breeze event lists, energy, edit stats, items_v4.synthetic.json)
metadata/round2/deflect_turns.synthetic.json	deflecting-turn labels used by deflect_w
metadata/fullrun/	episodes.tsv (id, split), its sha256, episodes/manifest.jsonl (tracks, mix, views, word timings, bad-directive windows), stats, selection report, root hours, CharacterBench eval sheets
metadata/episodes_flags.tsv	per-episode split, generator version, type, voice-reference corpora and bank ids, licence inheritance, Fish/Maya flag
metadata/breeze/	Breeze render manifest (manifest.json, manifest.d/ per job), bench/quality, first ids
data/episodes/*.tar.gz	episodes/<id>.json, s2s-episode/1: scene, participants (full character sheets), config, turns, controller events (NOTES), guardrail probes, persona tests, homeostasis, summaries, QC, word-level timeline, frame labels, render record
data/views/*.tar.gz	views/<id>/<agent>.json, s2s-view/1: per-agent training view (memory, target track, input tracks, labels, controller, loss-mask notes)
data/views_retronoted_v1/*.tar.gz	earlier retro-noted views (same ids)
data/render_qc/*.tar.gz	render_qc/<id>/render_stats.json: stage timings, gender QC + voice bank ids, speaker drift / retakes, emotion QC
data/tokens/*.tar	tokens/<id>.npz: Mimi tokens, uint16 [32, T] at 12.5 Hz, keys track/P* (each speaker's own track), others/P* (everyone else mixed), mix/all. voice/P* keys removed (see below)
generation/v3/	sheets (6,701 sheets), scripts (accepted, pre-render), rejected, index/syn-*.jsonl (per-unit generation log: type, accepted, timings, sheet), run/targets.json
generation/breeze/errors-*.tar.gz	per-episode render error records
generation/code_snapshots-*.tar.gz	the exact code snapshots the workers ran (breeze/code/<commit>, v3/code/<commit>)
repo/	from CharacterPlex (see Reproduce): generation code models/personaplex/finetune/episodes/ (gen_v3.py, render.py, breeze_*.py, sheet.py, views.py, voice_bank.py, nv_bank.py, ...), harness/data_v4/, harness/fullft/data_audit/, training config charplex_v4.yaml, card_v2_golden.json, docs/frontend_schema/ (schema + genuine samples), docs data_v4.md, data.md, card_v2.md, data_audit.md, MODELS_AND_LICENSES.md, PROVENANCE.md
tools/	the scripts that built and uploaded this release

Files marked .synthetic are filtered copies: entries for AMI/ICSI (and any other human-corpus id) were removed, so their sha256 differs from the one pinned in MANIFEST.json. All other metadata files are byte-identical to the originals. Paths inside the JSON (e.g. npz, views, audio_dir) are the original cluster paths under /scratch4weeks/hg288ju/speech-systems/; map s2s-episodes/breeze/tokens/<id>.npz → tokens/<id>.npz in the token shards, and so on.

Token shards and voice prompts. The original npz files also contain voice/P* (60-frame Mimi encodings of the human voice-reference clip of each speaker). These are encodings of third-party recordings, so they are removed. The other arrays are bit-identical to the originals (checked by round-trip at build time). The views reference voice_prompt: voice/P*: to train with the v4 config, re-encode the voice prompts from the voice bank on cluster storage, or replace them with your own prompts.

Generation method

1. Sheets (sheet.py, gen_v3.py): a sampled role/setting/trait profile; Qwen3.5-122B writes a front-end character sheet (fabric-character-library-v1 schema, repo/docs/frontend_schema/). Trait values are sampled in code (the key names the high end) and are binding.
2. Scripts (gen_v3.py): one sheet → K=5 episodes of different types. Setup call, one writer call per segment (segments start where NOTES land), repair, backchannels, then a Gemma 4 31B judge. Cross-cutting: "are you an AI?" tests, jailbreak / prompt-leak probes, banned topics, CTV-0 emotion tags. Accepted → scripts/, else rejected/. Seeds are in each episode's provenance.seed; code commit in provenance.generator.git_commit / V3_CODE_ID.
3. Render (render.py, breeze_worker.py): voice chosen from a bank of real speakers (Seamless, EARS, VCTK, L2-ARCTIC, Expresso) with gender/F0 QC; Breeze TTS 2 renders each turn (emotion QC, speaker-similarity drift check, retakes); human non-verbal clips spliced where TTS cannot render them; Qwen3 forced aligner gives word times; tracks are mixed on the scripted timeline and encoded with Mimi (32 codebooks, 12.5 Hz).
4. Views (views.py): one training view per agent (target track, input tracks, memory, NOTES/controller, labels, loss masks).
5. v4 (harness/data_v4/, metadata/v4/code/): energy rule on Mimi-decoded tokens around every event label, event edits, behaviour-coverage weights, held-out check; frozen in MANIFEST.json.

Reproduce / rebuild

• Code: git@github.com:utkarshyadav009/CharacterPlex.git (private), commit noted in INVENTORY.json sources (CharacterPlex@<commit>). Upstream model code is cloned by scripts/clone_upstream.sh (Breeze TTS 58ec70c).
• Training config: repo/models/personaplex/finetune/trainer/configs/charplex_v4.yaml (sha256 pinned in MANIFEST.json). Its roots also include the human corpora, which you must prepare yourself.
• Human corpora: download and shard with data/s2s/download_seamless.py, download_seamless_lively.py, download_ami_icsi.sh, download_ami_shard.sh, download_icsi.sh, download_expresso_ears_nonverbal.sh, finalize_shards.py, validate_shards.py in the CharacterPlex repo (not copied here).
• Re-render from the scripts: breeze_render.sbatch / render.py with the voice bank (voice_bank.py, nv_bank.py). About 14 min of TTS plus 9 min of emotion QC per episode on one GPU.
• Rebuild this release: tools/build_release.py, tools/finalize_inventory.py, tools/upload_dataset.py.

What is NOT included (and where it lives)

On cluster storage /scratch4weeks/hg288ju/speech-systems/ (purged after 4 weeks of no access):

Excluded	Location	Size	Why
Rendered audio (mix / tracks / clips)	s2s-episodes/breeze/render/<id>/*.wav	~5.2 GB (most episodes keep only tokens)	voice clones of third-party speakers + spliced human clips; redistribution right not verified
Voice reference clips	s2s-episodes/breeze/render/<id>/voices/	~13.2 GB	third-party human recordings
Voice prompts in npz (voice/P*)	s2s-episodes/breeze/tokens/	small	encodings of those recordings
Voice bank, non-verbal clip bank	s2s-episodes/voicebank, voicebank_accents_qc	~2.0 GB	human clips + speaker metadata
Reverse-annotated real conversations	s2s-episodes/v3/views_real, v3/index/rev-*, v3/run/reverse_records.json	~0.7 GB	derived from Seamless / Expresso
Human-corpus sidecar entries	charplex_v4/event_edits_v4.json (219 AMI/ICSI keys), charplex_v4/work/*human*	~20 MB	annotations of AMI/ICSI
Render-time episode.json copies	s2s-episodes/breeze/render/<id>/episode.json	~3.9 GB	duplicates of episodes/<id>.json
Logs, claim markers	s2s-episodes/breeze/logs, claims	~1.3 GB	operational
Pilot / trial / test runs	s2s-episodes/pilot (17 GB), api_test, render_trial*, trial*, test, listen_v2, v22dev	~19 GB	superseded
Generator LLM weights	s2s-episodes/models	~132 GB	on HF upstream
Human corpora, raw	data/ (seamless 651 GB, ears 65, ami 56, expresso 36, dipco 23, icsi 20, l2arctic 7.6, vctk 6.1, nonverbaltts 4.0)	~874 GB	third-party recordings
Human corpora, Mimi shards / roots	finetune/shards32/{seamless,ami,icsi,expresso,ears}, stage0/data/roots/	~12 GB	third-party recordings

Human corpora used in v4 training (names and licences only): Seamless Interaction (CC BY-NC 4.0), Expresso (CC BY-NC 4.0), EARS (CC BY-NC 4.0), AMI (CC BY 4.0), ICSI (CC BY 4.0). Others on storage used for voices / clips / evaluation: VCTK (CC BY 4.0), L2-ARCTIC (CC BY-NC 4.0), VocalSound (CC BY-SA 4.0), NonverbalTTS (treat as NC), DiPCo (CDLA-Permissive). Fisher is not used.

Limitations

• Synthetic speech: TTS prosody, rendered laughs/chuckles mostly fail the loudness rule (dropped in v4); groan and sob unsupported; scream has only 91 usable labels. No new audio was produced for v4.
• English only; sheets and scripts reflect the writer LLM's style and biases (UK-leaning roles and settings).
• Known gaps from evaluation: the model holds facts but deflects instead of saying them; relationship fields and traits barely move output; eval cards say "Visitor" while training cards always name the user.
• 50 held-out character voices are reused as the scripted user voice.
• Without the human roots, charplex_v4.yaml cannot be run as is; the gradient shares (synthetic 34 %, human 63 %, acted 3 %) need the human corpora.
• The provenance.licence.voices text of some pilot episodes says "designed voices; no real person cloned", but their render records show a VCTK reference voice; treat every episode as a voice clone.