shahnameh-qwen3-4b-multilingual-lora

Shahnameh Qwen3-4B — Multilingual LoRA

A LoRA adapter that makes Qwen3-4B-Instruct-2507 answer in Shahnameh-style verse in Persian, English and German.

This is the most consistent of the three models: clean script output in all three languages, at the cost of weaker Persian form.

Measured quality

For the strongest Persian verse, use SaeedAngiz1/shahnameh-qwen3-4b-fa-lora, accepting occasional script contamination.

Example

Q: Is it better to speak or to stay silent?

Speak not for words' sake, nor without thought,
Nor when you know not what you say is worth.
For words like arrows fly from hand to hand;
One that falls false may wound a friend, and none.

English is fluent and genuinely verse-like. German is fluent but drifts off-domain.

Usage

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "SaeedAngiz1/shahnameh-qwen3-4b-multilingual-lora",
    max_seq_length=1024, load_in_4bit=True, dtype=None)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are a poet who answers in the voice of Ferdowsi's Shahnameh, in English verse."},
    {"role": "user", "content": "Who was Rostam?"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
ids = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to("cuda")
out = model.generate(**ids, max_new_tokens=256, do_sample=True,
                     temperature=0.7, top_p=0.8, top_k=20,
                     pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

Training

Qwen3-4B-Instruct-2507 (4-bit) · LoRA r=16, lora_alpha=16 · lr 2e-4 · 3 epochs · 270 examples across fa/en/de · response-only loss · 2× Tesla T4 · 411 seconds. Best validation loss 2.553.

Honest limitations

• Facts are frequently wrong. The dataset teaches voice, not Shahnameh genealogy. Expect invented names and inverted relationships.
• Persian has never been reviewed by a native speaker. Rhyme and script purity are measured mechanically; meter (عروض), grammar and classical register are unverified.
• Output varies run to run at temperature=0.7. Judge on averages, not favourites.
• Trained on only 270 (or 90) examples. That is enough for style, not for prosody.

How it was built

Trained on Kaggle's free tier: 2× NVIDIA Tesla T4. Total compute for the whole project is under three hours. Full method, measurements, charts, and a documented record of four separate failures: https://github.com/SaeedAngiz1/shahnameh-slm

Credits

Ferdowsi's Shahnameh (completed 1010 CE). The training verse is newly written in his style — it is not his text.