Why Persian at all?
English-centric open models have two known pains in Persian: tokenizer word-splitting and unfaithfulness to output format. Light fine-tuning of a Persian-friendly open model (Qwen or Gemma) on a few hundred samples fades both pains inside your domain — without full-training costs.
One-minute prerequisite: hardware and tools
QLoRA loads the base model at 4-bit: a 7–8B model at Q4_K_M needs about 5 GB of GPU RAM and, with gradient checkpointing and a small batch, fits on a common 24 GB card (RTX 3090/4090) or even a 12 GB one. Tools: PyTorch + bitsandbytes + PEFT + TRL (or LLaMA-Factory if you want a GUI).
Data: clean Persian, small but consistent
We target behavior, not new knowledge, so hundreds to a couple thousand samples suffice. Pattern: "your team's real question + an answer in exactly your desired format". Three rules: write texts with proper Persian spacing and no broken words (broken samples produce a broken model); vary formats (email, report, short answer, table); set ten percent aside for eval and never touch it.
Execution: ten Python lines that do the job
Load the 4-bit model with bitsandbytes, build LoraConfig (rank 16, alpha 32, zero dropout, target=all-linear), then SFTTrainer for two to three epochs at 1e-4. Runtime with two hundred samples is tens of minutes, not hours. Output: a few-hundred-MB adapter.
Evaluation: your own Persian test set
Measure before and after on that eval: format accuracy (did the table stay a table?), Persian integrity (broken words? half-English leftovers?) and preserved knowledge — LoRA does not change knowledge; if out-of-domain answers degrade, the problem is data volume or overfitting, not the tool.
Deployment: the adapter beside the model
vLLM serves it with --enable-lora, multiple adapters at once; Ollama accepts an ADAPTER in a Modfile. For an internal API that is all it takes. And remember: version the adapter like code — each dataset addition gets a new version, an eval edit and a short CHANGELOG.
