Yooz-Light-v3-Qwen3.5-0.8B

Yooz-Light v3

Fast dictation touch-up model for the Yooz suite. Qwen3.5 0.8B (6-bit g64 MLX, fused from the teacher-KD QAT base), fine-tuned to correct raw speech-to-text output: punctuation, capitalization, spoken-number conversion, grammar, and light cleanup, in a strict JSON contract.

This is the production fast tier. Its sibling Yooz-Quality-v3-Qwen3.5-4B is the quality tier for full rewrites.

Behavior contract

• Input: a system instruction plus the raw transcription as the user turn.
• Output: a single JSON object, {"result": "corrected text"}, nothing else.
• Non-thinking, always. The chat template hard-pins the closed <think> block that matches the fine-tune's training rendering. Consumers do not need to pass enable_thinking=False; the flag is ignored.
• Fused weights only; no LoRA adapters ship in this repo.
• generation_config.json lists both stop ids (<|im_end|>, <|endoftext|>); generation terminates correctly out of the box.

Evaluation

Held-out test split (3,624 dictation-correction samples, proofread + rewrite), greedy decoding, versus the 9B incumbent used before the KD program:

*Python mlx-lm harness on an M4 Pro, batch 1; production Swift latencies are lower.

Yooz-Light v3 matches the 9B reference on overall quality at a fraction of the resources and clearly beats it on proofread-style corrections.

Usage (mlx-lm)

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("YoozLabs/Yooz-Light-v3-Qwen3.5-0.8B")

prompt = tokenizer.apply_chat_template(
    [
        {
            "role": "system",
            "content": (
                "Fix grammar, punctuation, and capitalization in this "
                "voice transcription. Convert spoken numbers to digits. "
                'Return JSON: {"result": "corrected text"}'
            ),
        },
        {"role": "user", "content": "the meeting is tomorrow at three pm"},
    ],
    tokenize=False,
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=256,
               sampler=make_sampler(temp=0.0)))
# {"result": "The meeting is tomorrow at 3:00 PM."}

Greedy decoding (temp=0.0) is recommended; the model is trained for deterministic correction, not creative generation.

Lineage

• Base: YoozLabs/Qwen3.5-0.8B-qat-lean-4bit-mlx, a teacher-KD QAT 4-bit g64 MLX build that repairs the aggregation collapse of community 4-bit PTQ at this size.
• Fine-tuned (QLoRA, fused) on a proprietary gold-standard corpus of 18,285 real dictation transcriptions with two correction targets per entry (proofread and rewrite).
• Ships at 6-bit g64: re-quantizing the fused LoRA delta at 4-bit on the QAT-optimized grid erases part of the fine-tune gain; 6-bit preserves adapter parity (verified on the full held-out split).
• Supersedes the Yooz-Light v2 / v2.1 working models.

Limitations

• English-centric; trained on dictation/STT output, not general text.
• Training sequence window 512 tokens; intended for utterance- to paragraph-length corrections.
• The JSON contract is the supported interface; free-form prompting is out of scope.

Contact

[email protected]

Part of the Yooz ecosystem. Sovereign Intelligence.