Views
No views yet
| Metric | Score |
|---|---|
| Wild pass | 84.48% |
| Hard-negative | 5/5 |
| Smoke | 8/8 |
| p50 latency | 676 ms |
| Ship-form size | 750 MB |
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)| 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 | PRIMARY |
| Re-training / stacking | adapter |
1from transformers import AutoModelForCausalLM, AutoTokenizer
2import torch
3
4model = AutoModelForCausalLM.from_pretrained(
5 "VoiceScribe/qwen3-5-0.8b-dictation-corrector-cuda-int4-nf4",
6 load_in_4bit=True,
7 device_map="cuda",
8 trust_remote_code=True,
9)
10tokenizer = AutoTokenizer.from_pretrained("VoiceScribe/qwen3-5-0.8b-dictation-corrector-cuda-int4-nf4", trust_remote_code=True)
11
12messages = [
13 {"role": "system", "content": "Корректор русской диктовки. Убери слова-паразиты ..."},
14 {"role": "user", "content": "Запушил коммит в гитхаб репозиторий"},
15]
16prompt = tokenizer.apply_chat_template(
17 messages,
18 add_generation_prompt=True,
19 tokenize=False,
20 enable_thinking=False, # CRITICAL for Qwen3.5
21)
22inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
23out = model.generate(**inputs, max_new_tokens=200, do_sample=False)
24print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
25# Expected: "Запушил коммит в GitHub репозиторий"VoiceScribe/qwen3-5-0.8b-dictation-corrector-mlx-{bf16,8bit,4bit} (V15 R-3, 93.1% wild)1@software{voicescribe-corrector-v15r3-2026,
2 title = {Voice Scribe Russian Dictation Corrector (Qwen3.5-0.8B V15 R-3, CUDA Win port)},
3 author = {Sabynin, Andrey},
4 year = {2026},
5 url = {https://huggingface.co/VoiceScribe/qwen3-5-0.8b-dictation-corrector-cuda-int4-nf4}
6}