ndgold/Qwen3-0.6B-EasyLanguage-4bit

Qwen3-0.6B-EasyLanguage (4-bit, MLX)

LoRA fine-tune of Qwen/Qwen3-0.6B that rewrites live speech transcripts into easy-to-read registers across 29 language locales — German (Leichte Sprache), French (FALC), English (Easy/Plain English), Spanish (Lectura Fácil), Easy-to-Read Arabic (Inclusion Europe / Information for All), Letlæst (Inclusion Europe ETR), and more — for the Live Linguist on-device captioner. Each locale follows its own national / European Easy-to-Read or plain-language standard. Quantized to 4-bit for Apple-silicon inference via MLX.

It splits run-ons into short sentences, drops disfluencies, keeps names/numbers, and stays in the input language. Trained on a lean prompt (no few-shots) so the register is internalized — shorter prompts, lower live-caption latency.

Evaluation (held-out test set; SARI = simplification quality)

lang	SARI ft	SARI stock	chrF ft	LID ft	compliance ft
de	47.56	32.98	41.93	1.0	0.985
fr	57.97	33.31	51.47	1.0	0.99
es	59.45	36.57	54.68	1.0	0.99
en	60.99	32.46	55.03	1.0	0.975
ar	50.04	50.11	46.58	0.99	0.945
da	53.67	35.58	51.18	0.99	0.93
et	47.65	30.48	50.87	1.0	0.995
fi	49.9	35.2	55.87	1.0	0.975
hi	51.87	35.16	44.18	0.99	0.805
it	54.0	49.18	48.83	1.0	0.8557
ja	8.39	8.29	41.38	1.0	0.965
ko	49.23	43.77	39.56	1.0	0.975
nl	51.68	38.96	48.21	0.99	0.89
pt-BR	55.34	37.54	51.92	1.0	0.755
pt-PT	58.17	38.96	58.38	1.0	0.93
ru	51.09	41.13	49.39	1.0	0.93
sk	51.19	40.59	46.52	1.0	0.96
sv	53.62	40.16	54.36	0.99	0.965
tr	50.51	31.73	49.61	0.995	0.97
vi	60.23	41.3	58.21	1.0	0.685
zh-CN	9.0	8.5	51.54	0.995	0.935
uk	52.12	39.63	51.13	0.995	0.99
ro	51.7	40.75	50.94	1.0	0.94
bg	50.83	34.17	52.76	0.985	0.98
pl	50.56	36.93	48.21	1.0	0.97
hr	51.3	33.88	50.85	0.995	0.99
nb	51.82	29.97	54.57	0.99	0.97
hu	49.74	38.56	51.06	1.0	0.985
cs	49.41	39.47	46.51	0.995	0.985

Usage (MLX)

from mlx_lm import load, generate
model, tok = load("ndgold/Qwen3-0.6B-EasyLanguage-4bit")
msgs = [{"role":"system","content":"<framework system prompt>"},
        {"role":"user","content":"Original: <utterance>\nRewritten:"}]
p = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False, enable_thinking=False)
print(generate(model, tok, prompt=p, max_tokens=96))

Sources & licenses

• Base model: Qwen3 (Apache-2.0).
• German seed data: tum-nlp/German4All-Corpus (German Wikipedia, CC BY-SA).
• Synthetic pairs (all other languages + de augmentation): spoken→easy-language pairs generated by Claude (Anthropic) — Leichte Sprache (de), FALC (fr), Easy/Plain English (en), Lectura Fácil (es), and 25 further locales following each language's Easy-to-Read / plain-language standard (Inclusion Europe "Information for All", Selkokieli, Lättläst, やさしい日本語, ISO 24495-1, …). Every pair is filtered by a deterministic validator suite (per-language sentence-length caps, language-ID, fidelity anchoring, number preservation, anti-parroting).
• Intended for the Live Linguist on-device live-caption simplifier. Not a general chatbot.