NicoV2-Saber

Built with Qwen. Public release of the existing NicoV2-Saber streaming action-tagger weights, September 27, 2026. Start with v6/lora_tagger.pt. The designated rollback is v2/lora_tagger.pt.

These are custom LoRA adapters and classification heads for Qwen2.5-Omni-3B, not a standalone language model, a standard PEFT adapter, a TTS voice, or the proposed NicoV2-Nelly architecture. The upstream base-model weights are not included.

Files

Version	Weight file	Status
v6	v6/lora_tagger.pt	Selected historical demo checkpoint
v2	v2/lora_tagger.pt	Designated rollback
v1	v1/lora_tagger.pt	Historical; no explicit control classifier
v3	v3/lora_tagger.pt	Historical, not selected
v4	v4/checkpoint-epoch-001.pt	Incomplete experimental run; no final checkpoint; do not deploy
v4b	v4b/lora_tagger.pt	Rejected due to semantic training-label contamination; do not deploy
v5	v5/lora_tagger.pt	Historical clean restart, not selected

Every .pt file is a byte-for-byte copy of its preserved original. Each contains lora, heads, config, step, and opt (None in these exported snapshots). The original config retains historical training IDs and paths for provenance; it does not include the recordings or transcripts. The separate inference_config.json files expose the relevant architecture/label fields without requiring checkpoint deserialization. Per-version modification notices identify these fine-tuned/custom components. Intermediate epochs and optimizer-bearing training snapshots are intentionally not uploaded, except the single explicitly marked incomplete v4 snapshot used to represent that historical run.

MANIFEST.json records each public file's size and SHA-256, checkpoint source-path mapping and source hashes. recommended.json is the preserved historical local selection record: its file/report paths refer to the original source repository, not to this Hub layout. It is not a new benchmark run or a deployment safety claim.

Architecture and runtime boundary

The checkpoints use rank-16 LoRA (alpha 32, scale 2) on q/k/v/o projections through Thinker layer 18, plus heads reading 2,048-dimensional hidden states through a 256-dimensional shared MLP. Heads predict intent, commit probability, per-slot BIO tags and, from v2 onward, abandon/revise/hesitate controls. The stored mean/std normalization buffers are part of the checkpoint and must be loaded too.

The selected historical runtime consumes 16 kHz audio in 250 ms blocks. This is a block setting, not a claim that end-to-end tool execution or speech completes within 250 ms. Saber uses its existing harness for slot resolution, policy, tools and replies; these weights alone are not a full conversational executive assistant.

Download and inspect

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download("Rak505/NicoV2-Saber", "v6/lora_tagger.pt")
checkpoint = torch.load(path, map_location="cpu", weights_only=True)
assert checkpoint["opt"] is None
print(checkpoint["config"]["layer"])  # 18

Pin revision to a Hub commit for reproducibility and verify the manifest hashes. These legacy .pt files are PyTorch serialization, not safetensors; only load trusted, hash-verified artifacts. No trust_remote_code is required to inspect them.

For inference, use the original tagger.lora_tagger.LoraTagger implementation from the NicoV2/VoiceSynth source checkout, with demo on PYTHONPATH:

from tagger.lora_tagger import LoraTagger
tagger = LoraTagger(ckpt=path, device="cuda", block_ms=250, compact=True)

That implementation requires the base model already cached locally and the source project's dependencies. The source repository is DerekWWang/voicesynth; access may be required. This weights-only publication does not change that repository's visibility or publish a standalone inference package. The repository revision observed during release was 08222e3f158e3870cfee6b97766fa2bc6103df85; it is not claimed to be the original training-code revision.

Evaluation and limitations

The existing v6 selection prioritized cancellation/negative safety. Its recorded real-voice development result was 19/25 resolved calls and 9/25 strict all-argument matches. v2 had 20/25 resolved calls on that narrow set. These are historical, repeatedly inspected development results, not unbiased held-out test estimates. Replay timing excludes synchronous ASR waiting and TTS. See recommended.json for the original scope and additional control/broad-synthetic results; no evaluation was rerun or checkpoint promoted as part of this publication.

Known limitations include negative-action errors, early commits, weak implicit-note and scheduling cases, and incomplete generalization to multi-frame requests and new tool schemas. None of these models is cleared for unsupervised external writes. Use a stateful sandbox, enforce authorization in the harness, and require appropriate confirmation before consequential actions. v4/v4b must not be chosen merely for a lower training loss.

License and attribution

The upstream Qwen2.5-Omni-3B model uses the Qwen Research License Agreement, not Apache 2.0. The copied LICENSE and Notice accompany this release. Its grant is for non-commercial research/evaluation; commercial use requires a separate license from Alibaba Cloud. Public download availability does not grant unrestricted commercial rights. No additional permissive license is assigned to the custom components by this upload.

This publication contains no API keys, raw audio, synthetic dataset, base-model weights or Nelly development checkpoints. Original local Saber files are unchanged.