Teaches two things the model gets wrong: realising the vocal burst that was actually asked for (a scream where a scream was written, not a laugh and not silence), and stopping when the line is finished instead of mumbling on past it.
Unlike the earlier quality adapters in this line, this one is not a null result. On 317 held-out preference pairs, pooled accuracy goes 0.707 → 0.943, and every one of the eight task families improves.
⚠️ Do not merge this adapter into the base weights#
merge_and_unload(), merge_adapter(), or any offline "bake the LoRA into the checkpoint" script will destroy the model irrecoverably.
This adapter targets audio_lm_heads.0 … audio_lm_heads.11 and text_lm_head — 12 of its 23 target modules. Those output heads are weight-tied to the input embeddings: tie_weights() makes audio_lm_heads[i].weight and audio_embeddings[i].weightthe same tensor, one allocation with two names. A merge adds B @ A · (α/r) into the head and therefore writes the same delta into the embedding table. The model then reads its own inputs through a matrix shifted by an output-side correction. It does not fail loudly — it degrades into noise, inside the checkpoint you just saved, with the original values gone.
Verify it yourself:
m = base.model if hasattr(base, "model") else base
print(m.audio_lm_heads[0].weight.data_ptr() == m.audio_embeddings[0].weight.data_ptr())
# True -> same storage
Load with PEFT and leave it unmerged. This is acoustically identical to a merge: an unmerged LoRA computes Wx + (B @ A)x · (α/r), a merged one (W + B @ A · (α/r))x — the same arithmetic in a different order. You give up a little inference speed and keep adjustable strength, stackability and the ability to switch it off.
14,932 preference pairs. Each pair is the same recording twice — one version the model should prefer, one it should not — so it cannot learn "prefer a different voice" or "prefer a calmer delivery". Four arms:
Vocal bursts (30 %).Chosen = the original clip with its burst intact. Rejected, in three kinds:
kind
share of the burst arm
what the negative is
swap
39.5 %
a different burst class from the same speaker, similar length — a laugh where a scream belongs
silence
35.1 %
the burst replaced by silence, faded in and out so total duration is unchanged
excise
25.4 %
the burst cut out entirely (duration shortens)
The swap kind is the point. With silence-only negatives the model can cheat: it only has to detect whether a burst is present, never which one. Making swap the largest kind forces it to identify the class.
Stopping (25 %). Three ways of teaching the model to fall silent:
truncated prompt — the script is cut a word or two early and the audio is cut and faded at the same point; rejected is the original, which keeps talking past the prompt.
two-sentence fade — only the first sentence is in the prompt and the second is faded to silence in the chosen audio. The stated duration is the FULL original length, so the model learns to say the first sentence at its natural pace and then produce silence — not to stretch it across the whole budget. Verified: 410/410 renders have exactly one speech segment whose sum equals the full original duration.
appended noise — chosen is the clean original; rejected has a random fragment from another take by the same speaker appended, so the negative is "kept going" rather than "was cut".
Quality (25 %) and speed (20 %) are carried over from the earlier corpora as controls, so the run can show it is not damaging what already worked.
Scored forward-only through every kept checkpoint on the same 317 held-out pairs — one record per pair per checkpoint, 5,706 records. That makes every comparison paired, with each pair its own control, rather than a marginal accuracy on a few dozen pairs.
Improvising past the end of a line was not merely unlearned — it was learned backwards. The starting adapter scores 0.258 on the two-sentence stop task: 8 right, 23 wrong. It actively prefers the take that keeps talking. After 112 steps it is at 0.871, and 1.000 by the end.
This would have been invisible in the pooled number: 31 pairs out of 317 moving from 0.258 to 1.000 shifts the pooled accuracy by 0.08, inside its own noise. Only the per-family paired scoring found it.
The two arms learn on completely different schedules#
All three stop families are finished by step 280. burst_silence is still below its starting value at step 224 (0.481 vs 0.407), first becomes significant at step 336, and peaks at step 784. A checkpoint chosen on the stop arm alone would be chosen roughly 500 steps too early.
burst_excise is the family this corpus did not solve (0.458 → 0.667). By construction: it is the only burst kind whose two sides differ in length, so it overlaps the speed-stretch task and the model can decide it on duration instead of on the burst.
step336 — the first checkpoint where all eight families are at or above the starting adapter on the paired logit and the burst arm has begun to move.
step896 — the largest paired gains in the ladder and the best pooled accuracy (0.9432).
Deliberately not best-by-val_loss (step 840) and not best-by-pooled-accuracy (step 616). step902 is a 6-step tail and the only late reversal in the run — use 896, not final.
Nobody has listened to any of these yet. Every number above is preference accuracy on held-out pairs. A listening test should compare four conditions — the starting adapter (ratelora_dpo_mix_r16/step245) plus step336, step616 and step896 — with two separate judgements per clip, "is this the burst that was asked for" and "does it stop", kept apart from any audio-quality rating.
In LoRA mode the trainer's reference policy is peft.disable_adapter(), which disables all adapters — so rew_chosen is measured against bare SFT3, not against this run's starting point. It was flat (3.29–3.49) across all 902 steps while rew_rejected fell from +0.295 to −3.87, so all margin came from pushing the rejected side down. Read rew_chosen > 0 as "better than bare SFT3", which is a weak guardrail, not a strong one.
import torch
from transformers import AutoConfig, AutoModel, AutoProcessor
from peft import PeftModel
BASE = "laion/moss-tts-local-transformer-4.55b-voice-acting-v2-sft3"
dev = "cuda:0"
proc = AutoProcessor.from_pretrained(BASE, trust_remote_code=True)
proc.audio_tokenizer = proc.audio_tokenizer.to(dev).eval()
cfg = AutoConfig.from_pretrained(BASE, trust_remote_code=True)
base = AutoModel.from_pretrained(BASE, trust_remote_code=True,
dtype=torch.bfloat16,
attn_implementation="sdpa").to(dev).eval()
model = PeftModel.from_pretrained(base, "laion/moss-va-sft3-burst-stop-dpo-lora",
adapter_name="bs").to(dev).eval()
# do NOT call model.merge_and_unload()
batch = proc([[{"role": "user", "content": prompt, "audio_codes_list": []}]],
mode="generation")
out = model.generate(input_ids=batch["input_ids"].to(dev),
attention_mask=batch["attention_mask"].to(dev).to(torch.bool),
max_new_tokens=340, do_sample=True,
temperature=1.0, top_p=0.95, top_k=25)
A different checkpoint:
model = PeftModel.from_pretrained(base, "laion/moss-va-sft3-burst-stop-dpo-lora",
subfolder="checkpoints/step336", adapter_name="bs")
Adapter strength (base scaling is α/r = 2.0; all measurements used w = 1.0):
def set_weight(model, name, w):
for module in model.modules():
sc = getattr(module, "scaling", None)
if isinstance(sc, dict) and name in sc:
if not hasattr(module, "_base_scaling"):
module._base_scaling = {}
module._base_scaling.setdefault(name, sc[name])
sc[name] = module._base_scaling[name] * float(w)
set_weight(model, "bs", 1.0)
Trained under prompt format hash 090c4ca315519a57 (inherited through the adapter chain). A duration-only script is byte-identical to the project's standard 073aeb09dc923376, so ordinary prompts work unchanged.
checkpoints/step{112,336,616,896} — the four reported above. The root copy is step896. All 18 checkpoints exist internally; these are the ones with published numbers.
Per-burst-class adapters: moss-va-sft3-vocal-burst-lora-adapters — note that this DPO adapter and those adapters attack the problem at different points and have not been tested together
onnx-community 
mlx-community