VoiceScribe/qwen3-5-0.8b-dictation-corrector-lora-adapter

Voice Scribe Russian Dictation Corrector (Qwen3.5-0.8B, V15 R-3, adapter)

PEFT LoRA adapter only (~5 MB). Apply on top of Qwen/Qwen3.5-0.8B via peft.PeftModel.from_pretrained. For community re-training or stacking.

Eval results (held-out wild_eval, 58 prompts × 9 sectors)

Metric	Score
Wild pass	96.55%
Hard-negative	5/5
Smoke	7/8
p50 latency	574 ms
Ship-form size	5 MB

Comparison:

• macOS V15 R-3 reference: 93.1% wild
• V14 baseline: 86.2%
• Qwen3-4B Q5 production (pre-LoRA): 48%
• This model: 96.55% (+10.3pp vs V14 baseline)

Training recipe (V15 R-3)

Mirrors macOS configs/r4_v15_extended.yaml byte-for-byte logical-recipe.

base               = Qwen/Qwen3.5-0.8B (vanilla, NOT Instruct)
LoRA rank          = 16
LoRA alpha         = 80 (rsLoRA mode -> effective scale 20.0)
target_modules     = q_proj, k_proj, v_proj, o_proj
layers_to_transform = last 16 of 24 (range(8, 24))
mask_prompt        = ON (assistant_masks via patched chat_template {% generation %})
max_steps          = 1100
batch_size         = 2
max_seq_length     = 384
lr_schedule        = cosine, peak 3e-5, warmup 100
weight_decay       = 0.01
optim              = adamw_torch_fused
precision          = bf16
seed               = 20260515
trainable params   = 720,896 (0.0957% of 753M)
data               = 1104 rows = V14 seeds (691) + V15 brand expansion (271) + V15 R-3 patches (142)

Intended use

• Russian dictation cleanup after ASR (GigaAM, Whisper, Parakeet)
• Conservative editing policy: remove filler (эм/ну/типа/короче), normalize Cyrillic IT terms (гитхаб -> GitHub), preserve all meaning
• NOT for general text editing, English text, creative writing, summarization, translation

Limitations

• 58-row eval set has ±1.72pp single-row noise
• Cyrillic <-> Latin choice on ambiguous brand spellings is judgment call (model may differ from expected byte-match)
• Trained on synthetic data only; real production telemetry collection planned for V16

Hardware ship matrix

Hardware	Recommended ship-form	This model?
RTX 5090 / 4090 24GB+	bf16
RTX 4070 / 4060 / 3060 8-16GB	INT8
RTX 2060 / 3050 / 4060 6-8GB	INT4 NF4
Re-training / stacking	adapter	PRIMARY

Inference

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "VoiceScribe/qwen3-5-0.8b-dictation-corrector-lora-adapter",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("VoiceScribe/qwen3-5-0.8b-dictation-corrector-lora-adapter", trust_remote_code=True)

messages = [
    {"role": "system", "content": "Корректор русской диктовки. Убери слова-паразиты ..."},
    {"role": "user", "content": "Запушил коммит в гитхаб репозиторий"},
]
prompt = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=False,
    enable_thinking=False,  # CRITICAL for Qwen3.5
)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
# Expected: "Запушил коммит в GitHub репозиторий"

Cross-platform variants

• macOS MLX: VoiceScribe/qwen3-5-0.8b-dictation-corrector-mlx-{bf16,8bit,4bit} (V15 R-3, 93.1% wild)
• CUDA bf16/INT8/INT4-NF4: this family (V15 R-3 Win port, 84.48-96.55% wild)
• OpenVINO: planned (separate venv for export; tracker WP#920)
• TensorRT-RTX W4A16: deferred (DeltaNet ONNX export blocked on Win-Py3.13-cu128 in 2026-05)

Citation

@software{voicescribe-corrector-v15r3-2026,
  title = {Voice Scribe Russian Dictation Corrector (Qwen3.5-0.8B V15 R-3, CUDA Win port)},
  author = {Sabynin, Andrey},
  year = {2026},
  url = {https://huggingface.co/VoiceScribe/qwen3-5-0.8b-dictation-corrector-lora-adapter}
}

Trackers

• macOS R&D: OpenProject WP#917 (V14), WP#919 (V15 R-3 macOS)
• Windows port: OpenProject WP#920 (this effort, achieved 96.55% vs macOS 93.1%)