utkarshUY/CharacterPlex-FullFT-v4

CharacterPlex FullFT v4 (step 2,765)

Private, research use only. A full-duplex speech-to-speech character model: a full-parameter fine-tune of NVIDIA PersonaPlex 7B (Moshi architecture, 8 agent codebooks, Mimi codec frozen). It plays a character from the front-end character sheet (fabric-character-library-v1, rendered as a v2 card) and follows narrative beats and live director notes.

This is the PRIMARY CharacterPlex checkpoint. The fallback is utkarshUY/CharacterPlex-LoRA-v1 (LoRA step 2,500). Code, serving and deployment: the private GitHub repo utkarshyadav009/CharacterPlex (docs/DEPLOY.md).

Files

File	What
model.safetensors	merged bf16 weights, 475 tensors, 8,371,408,896 params (same names/shapes as the base)
memory.json	pinned-memory layout the model was trained with (NOTES 64 frames, card ≤ 1000 tokens, memory_trained: true); the server reads it from the weights' folder
codebooks.json	8 agent / 8 user codebooks, delays
tokenizer_spm_32k_3.model	text tokenizer (from the base repo, unchanged)
tokenizer-e351c8d8-checkpoint125.safetensors	Mimi audio codec (from the base repo, unchanged, frozen in training)
base_config.json	the base repo's config.json
export_check.json, consolidate.json	export verification: merged file vs training model, max abs logit diff 0.0, argmax agreement 1.0
SHA256SUMS	SHA-256 of every file

Provenance

• Base: nvidia/personaplex-7b-v1 @ fdaf4090a61cb315c138a1faee287ffd6c716309.
• Training: Slurm job 47358 (8 × H200), 4,609 steps, full-parameter SFT with ZeRO-2; LR temporal 2e-6, depformer 4e-6, embeddings/head 2e-6; global batch 64; cosine, warm-up 150. 7.889 B of 8.371 B params updated (depformer steps 8–15, the user-audio path, inactive). Code experimental/charplex-run @ 120dc17 (+ stop-rule fix bdd1ddb).
• Data: v4 manifest sha256 262e0c97bdf3d6aba985eb636d047319cfceda064977525c073f3f68a114bf8a (see the dataset repo utkarshUY/CharacterPlex-Synthetic-Data-v4); gradient share ~63 % natural human dialogue, ~34 % synthetic, ~3 % acted.
• Selection: step 2,765 chosen by the predeclared rule (weights obedience 0.30, persona 0.20, recall 0.15, talk-over 0.15, latency 0.10, UTMOS 0.10, identity floor), not the final step.

Conditioning: v2 card (required)

Use card format v2 (trainer/card_v2.py, 700-token budget) compiled from the real front-end JSON via orchestrator/fabric_adapter.py. Trait polarity is converted (front-end value → 100 − v). The model expects pinned memory segments CARD / SCENE / ROSTER / NOTES; live director notes, user facts and beat lines go into NOTES (set_notes, ~16 ms, no reconnect). v1 cards work but lose most of the director-control gain.

Results (frozen CharacterBench v2.1, matched protocol)

	base	LoRA 2,500	this
Persona (P+L)	5.10	5.36	5.50
Identity under a conflicting note	8.12	7.75	7.42
Director-note lift over control (pts)	0	38.9	65.9
Arc / memory reveal on cue	1.9 %	28.8 %	92.3 %
User's name recalled (no runtime memory)	69 %	38 %	19 %
Still talking at user's turn end (no floor gate)	6.5 %	21.3 %	17.1 %
Reply latency	1.20 s	1.24 s	1.29 s
UTMOS	3.06	2.94	2.91
On-cue vocal events	1 %	8 %	19 %
10-min persona, first → last 3 min	5.25 → 3.50	5.33 → 3.83	5.92 → 6.08

With the CharacterPlex runtime (not part of the weights):

• User-fact memory (userMemory): name recall 17 → 60 %, pet 6 → 67 %, job 50 → 65 % (CB M suite, 52 paired).
• Floor gate (floor: {"mode": "staged", "hold": true}): talk-over at user turn end 18.8 → 7.3 %, latency +0.08 s.

Limitations

• Recall without the runtime memory is poor: the model hears and holds facts but often deflects instead of saying them.
• Talks over the user far more than base unless the runtime floor gate is on.
• Identity under conflicting director notes is lower than base (7.42 vs 8.12).
• Individual traits move behaviour little; the card as a whole does.
• Vocal events (laugh, whisper, sigh) appear on cue only ~19 % of the time; crying and screaming are unconfirmed; it does not react to the user's own laughs/sighs; gasps/long pauses sometimes fire in neutral talk.
• Context window ~194 s; long-range facts need the runtime memory.
• English only.

Licences

• Weights: derivative of NVIDIA PersonaPlex 7B — NVIDIA Open Model License; Mimi codec and Moshi architecture from Kyutai (CC BY 4.0). Research use; keep this repo private.
• Training data include third-party conversational corpora with their own (some non-commercial) licences; see the dataset card. Do not use commercially.