qwen3-4b-prosody-rag-lora-mlx

Qwen3-4B LoRA — Prosody & Phonology (RAG-grounded QA)

LoRA adapter for mlx-community/Qwen3-4B-Instruct-2507-4bit, fine-tuned to answer questions about prosody, intonational phonology and phonetics (with a focus on Italian and Italo-Romance varieties) in a retrieval-augmented regime: the model expects one or more source excerpts in the user message and produces grounded, terminologically precise answers in English or Italian.

Trained with mlx-lm on an Apple M4 Pro.

Intended use

This adapter is designed as the generation head of a RAG system over a linguistics paper collection. Prompt format used in training:

system: You are an expert assistant in prosody and phonology. Answer the
        question based on the provided excerpt, using precise terminology.
user:   ESTRATTO (<source>, p. <n>):
        <excerpt text>

        DOMANDA: <question>

Without excerpts in the prompt, the model falls back to generic behaviour — it was deliberately not trained to answer from parametric memory, to avoid hallucinated bibliography.

Usage (MLX, Apple Silicon)

from huggingface_hub import snapshot_download
from mlx_lm import load, generate

adapter = snapshot_download("GiusMagi/qwen3-4b-prosody-rag-lora-mlx")
model, tokenizer = load("mlx-community/Qwen3-4B-Instruct-2507-4bit",
                        adapter_path=adapter)

messages = [
    {"role": "system", "content": "You are an expert assistant in prosody and phonology. Answer the question based on the provided excerpt, using precise terminology."},
    {"role": "user", "content": "ESTRATTO (...): <excerpt>\n\nDOMANDA: <question>"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False,
                                       add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=400))

GGUF (CPU / cross-platform)

prosody-rag-4b.Q4_K_M.gguf (2.5 GB) is the adapter fused into the base model and quantized Q4_K_M — runs anywhere via llama.cpp, LM Studio or Ollama:

llama-cli -m prosody-rag-4b.Q4_K_M.gguf --temp 0.3 -cnv
# oppure, con Ollama:
#   FROM ./prosody-rag-4b.Q4_K_M.gguf  → ollama create prosody-rag

Training details

Limitations

• Answers should always be checked against the retrieved sources; the adapter improves terminology and grounding style, it does not guarantee factuality.
• Domain-narrow: outside prosody/phonology it behaves like the base model.
• The training corpus is not distributed with this adapter.