laion/moss-va-sft3-quality-dpo-lora

MOSS VA SFT3 — audio-quality DPO LoRA (rank 16)

<!-- moss-tied-heads-warning -->

⚠️ Do not merge this adapter into the base weights

merge_and_unload(), merge_adapter(), and any offline "bake the LoRA into the checkpoint" script will destroy the model irrecoverably. This is not a performance caveat. Read this before you write a deployment script.

Why

This adapter targets audio_lm_heads.0 … audio_lm_heads.11 and text_lm_head — 12 of its 23 target modules. In this architecture those output heads are weight-tied to the input embeddings: tie_weights() sets

audio_lm_heads[i].weight  IS  audio_embeddings[i].weight     # the same tensor, not a copy
text_lm_head.weight       IS  transformer.embed_tokens.weight

They are one allocation with two names. So when a merge adds B @ A * (alpha/r) into the head weight, it writes that delta straight into the embedding table at the same time. The model then reads its own inputs through a matrix that has been shifted by an output-side correction. Generation does not fail loudly — it degrades into noise or into a fixed babble, and the damage is inside the checkpoint you just saved. There is nothing to unmerge afterwards, because the original values are gone.

Verify it yourself in three lines

Do not take our word for it:

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, merging corrupts the embeddings

What to do instead

Load with PEFT and leave the adapter unmerged. Set its strength through the scaling factor:

from peft import PeftModel

model = PeftModel.from_pretrained(base, "<this repo>", adapter_name="a").to(dev).eval()
# do NOT call model.merge_and_unload()

def set_weight(model, name, w):
    """Scale one named adapter's contribution.  alpha/r is its own base scaling."""
    for module in model.modules():
        scaling = getattr(module, "scaling", None)
        if isinstance(scaling, dict) and name in scaling:
            if not hasattr(module, "_base_scaling"):
                module._base_scaling = {}
            module._base_scaling.setdefault(name, scaling[name])
            scaling[name] = module._base_scaling[name] * float(w)

set_weight(model, "a", 1.0)
model.base_model.set_adapter(["a"])      # several adapters can be active at once

This sounds identical to a merge. An unmerged LoRA computes Wx + (B @ A)x * (alpha/r), which is exactly what the merged weight W + B @ A * (alpha/r) would compute — the same arithmetic, in a different order. You give up a small amount of inference speed and you keep the ability to change the weight, stack several adapters, or turn one off. Nothing about the sound changes.

If you are stacking adapters

Set each one's scaling separately and activate them together with model.base_model.set_adapter([...]). Note that stacking is not free: in our own measurements a deep stack held audio quality but destroyed intelligibility (word error 0.063 → 0.554). Add adapters deliberately and measure.

If you maintain code that merges

A regex over module names is not enough — the reliable test is identity of storage. Group the modules by weight.data_ptr() and refuse to merge into any group with more than one member. lora_bank.py in LAION-AI/Humaneness-Voice-Demo-Server does this and asserts on the merge path.

Preference-trained to prefer cleaner recordings of the same performance.

Read this first: in our own evaluation this adapter does not improve generated audio quality. Across 16 prompts, paired against the base model, the best checkpoint scores +0.017 DNSMOS (t +0.45) — indistinguishable from no adapter — and the most heavily trained checkpoint is −0.119 (t −2.14), i.e. slightly worse. We publish it with the numbers rather than without them; the null result is the finding.

What it was trained to do, in plain language

We had 24,379 real recordings that were run through a speech enhancer (SIDON). For each one we kept the pair (enhanced take, original take) — the same performance, the same speaker, the same words, one cleaner than the other. Then we used Direct Preference Optimization to nudge the model toward producing audio that looks like the cleaner half of each pair.

The point of pairing the same take against itself is that the model cannot learn "prefer a different voice" or "prefer a calmer delivery" — those are held constant. The only thing that varies is signal cleanliness.

Three guardrails decided which pairs were allowed in:

• —the enhanced version had to score at least +0.10 DNSMOS better (mean gain +0.253, max +1.927);
• —it had to stay closer to the same speaker than a same-voice baseline (mean cosine margin 0.490), so quality could not be bought by changing the voice;
• —its emotion embedding had to stay close to the original, so the performance could not be ironed flat.

Honest limitation of the data

All 24,379 pairs come from one enhancer (enhanced_variant = raw_sidon). A second arm using voice conversion did not survive the gates. So this adapter learned one tool's notion of "clean", not a general one.

Results

DNSMOS is the right primary metric here because it is what selected the training data. If the adapter learned the objective, generated audio should score higher. It does not.

16 prompts × 3 samples per checkpoint, all conditions generated from the same prompts under the same seed, samples averaged within a prompt first, so n = 16 and the t is over prompts.

With ten conditions tested at n = 16, |t| ≈ 2.1 is roughly what chance produces, so the two apparently significant negatives should be read as "suspicious", not "established". The defensible summary is: no checkpoint is better than the base model, and the extremes look worse.

One counter-signal worth naming: step1504 has the highest genuineness of any condition (+0.197, t +1.90) while having the worst DNSMOS. If that survives a listening test it would matter, because genuineness is closer to what a voice-acting model is for than DNSMOS is. It has not been listening-tested.

Why the recommended checkpoint is an early one

The preference task is solved by step 376: accuracy hits 1.0000 and validation loss goes flat. Everything after that inflates the margin without learning anything.

reward(chosen) more than doubles from step 376 to the end at constant accuracy — about 1.0 nat per token of drift away from the supervised reference model, for nothing.

A warning for anyone selecting checkpoints by validation loss: for DPO a lower validation loss just means a larger margin, so that criterion prefers the most aggressive checkpoint by construction. It is not a quality criterion. Every checkpoint here was preserved specifically because the trainer's own pruning would have discarded the early ones.

How it was trained

Inference

⚠️ Never merge this adapter into the base weights. Its target modules include audio_lm_heads.0 … audio_lm_heads.11, whose tensors share storage with audio_embeddings.N.weight (weight tying). Merging corrupts the embedding table irrecoverably.

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-quality-dpo-lora",
                                  adapter_name="qual").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=50)

Pick a different checkpoint with subfolder:

model = PeftModel.from_pretrained(base, "laion/moss-va-sft3-quality-dpo-lora",
                                  subfolder="checkpoints/step1504",
                                  adapter_name="qual")

Adapter strength (base scaling is α/r = 32/16 = 2.0; all measurements used w = 1.0):

for module in model.modules():
    scaling = getattr(module, "scaling", None)
    if isinstance(scaling, dict) and "qual" in scaling:
        module.scaling["qual"] = 2.0 * w

Prompt format

Trained under prompt format hash 090c4ca315519a57 (inherited from the rate stage). A duration-only script is byte-identical to the project's standard 073aeb09dc923376, so ordinary prompts work unchanged.

Checkpoints

checkpoints/step{188,376,752,1128,1504} — exactly the five that were evaluated above. Nothing is published here without numbers. The root copy is step376.

Related

• —Base model: `laion/moss-tts-local-transformer-4.55b-voice-acting-v2-sft3`
• —Initialised from: `laion/moss-va-sft3-rate-lora`
• —Sibling trained on a 75/25 quality+speed mix: `laion/moss-va-sft3-quality-speed-dpo-lora`
• —General-purpose DPO adapter: `laion/moss-va-sft3-dpo-lora-p2`