jplstar

Telisk 1.0 — JPLSTAR

Fine-tune of Qwen/Qwen2.5-7B-Instruct for Standard Persian → Abizi.
Qaeni is treated as a closely related South Khorasan variety in the training mix, not as a separately scored test language.

Trained with Unsloth QLoRA on 2× NVIDIA T4 (Kaggle). Adapter merged. Repo: amirhosseinjpl/jplstar.

This is a dialect transfer model. It is not a general chatbot. Without the system turn used at train time, the base Qwen identity can reappear. That is expected.

Inference (matches eval)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "amirhosseinjpl/jplstar"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype=torch.bfloat16, device_map="auto"
)

messages = [
    {
        "role": "system",
        "content": "Translate the following Standard Persian sentence into Abizi dialect.",
    },
    {"role": "user", "content": "دیشب رفتم مغازه."},
]

prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=False,
    pad_token_id=tok.eos_token_id,
)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Greedy decoding is what produced the numbers below. Sampling is for the demo Space only.

Splits

Held-out rows were isolated from train-Final. No reverse-pair cut from the same sentence into eval.

Eval

Metric: character overlap (F1 over character bags after yeh/kaf normalize). Exact match is the wrong headline: Abizi spelling is not stable.

Valid is a 30-row file. Do not cite a 50-row file as valid. Test > valid on this metric is compatible with split difficulty and small-n noise. It is not proof of generalization. Human gold-vs-pred is the judge. Overlap is a proxy. Not BLEU.

Train stack

• Base: Qwen2.5-7B-Instruct
• Method: Unsloth QLoRA → merge
• Hardware: 2× NVIDIA T4
• Format: ChatML / instruction routing

Hyperparameters live in the training notebook, not in this card.

Files

• Merged weights in this repo
• Demo Space: amirhosseinjpl/telisk (loads this model)

License

Apache-2.0 for the fine-tune artifacts, subject to the Qwen2.5-Instruct license of the base.