ibnsina-30b

ibnsina-30b — Persian continued pretraining of Qwen3-30B-A3B-Base

⚠️ این یک مدل پایه است، نه یک دستیار گفت‌وگو. فقط متن را ادامه می‌دهد؛ دستور نمی‌گیرد، نقش‌بازی نمی‌کند و برای مکالمه تنظیم نشده است. برای گفت‌وگو به فارسی از ابن‌سینا ۳B استفاده کنید. این مدل برای پژوهشگران و توسعه‌دهندگانی است که می‌خواهند روی آن fine-tune کنند.

⚠️ This is a BASE model, not a chat assistant. It continues text; it does not follow instructions, hold a conversation, or refuse anything — there is no instruction tuning and no chat template. For Persian conversation use IbnSina-3B. This model is for researchers and developers who want to fine-tune on top of it.

IbnSina-30B is Qwen3-30B-A3B-Base continued-pretrained on 6.0 B tokens of the IbnSina Persian mix. It is the third and largest model of the IbnSina family and the only one that is not trained from scratch: the 1.5B and 3B are Persian-first models built from zero, while this one adapts a strong open Mixture-of-Experts base to Persian. 30.5 B total parameters with 3.3 B active per token, so it runs at roughly 3B-model speed while holding 30B-model knowledge. Author: Sina Meraji · ORCID 0009-0002-8028-1932 · github.com/ibnsina-llm.

Files

ollama run hf.co/ibnsina-llm/ibnsina-30b          # completes text, does not chat
llama-cli -m ibnsina-30b-Q4_K_M.gguf -p "ایران کشوری در غرب آسیا است که"

Model

Evaluation

Log-likelihood multiple-choice scoring, the same harness and prompts as the rest of the family, with the untouched base model as the control — the honest question for a continued-pretraining run is what did Persian training change, not how does it rank.

What 6 B tokens of Persian did: the held-out Persian bits-per-byte fell by 14 % (0.5646 → 0.4835; perplexity per token 3.77 → 3.12), the clearest signal and the one the training optimised. On the knowledge side PersianMedQA rose 5.2 points (51.4 → 56.6 %, gains in most fields) and ParsiNLU multiple choice 1 point; paraphrase detection moved off the base's constant-answer prior (50.0 → 56.7 %). Entailment fell 4.8 points (58.3 → 53.5 %): the base leaned on one label (1,139 of 1,673 answers were "entailment") and was often right; ours spreads its answers across the three labels and gets fewer of them — a calibration change more than a knowledge loss, but a regression on that row and reported as one. Under the generative protocol (answer with the option number, temperature 0) the two are level on all 5,235 questions (57.0 vs 56.8 %) because ours fails the number-only format eight times more often (411 vs 52 unparsed replies); on the questions it does answer it is 4.5 points ahead (61.8 vs 57.3 %). The base model's English and general abilities were not re-measured; a 0.02 % extra pass of mostly Persian data is unlikely to move them much, but that is an expectation, not a measurement.

Intended use and limitations

A base model for research and for building Persian systems: fine-tune it, probe it, or use it for completion. It has no instruction tuning, no chat template, no safety tuning of its own beyond what the base model carries, and no behaviour policy — the scope banner and the symmetric-respect policy of the IbnSina chat models do not apply here, because nothing in this training taught them. It inherits the base model's knowledge cutoff and its biases, plus whatever the Persian mix added. Knowledge questions, medical, legal and financial matters: not a reliable source, and with no refusal behaviour it will answer anything. Verify everything that matters.

Licence

Apache-2.0, the same as the base model. Training-data licences are per source (see the repo's licence table); the mix admits only sources whose licences permit an Apache-2.0 release, and no source text is redistributed.

Citation

@software{ibnsina2026, title={IbnSina: an open Persian-first language model family}, author={Meraji, Sina}, year={2026}, url={https://github.com/ibnsina-llm}, note={ORCID 0009-0002-8028-1932}}

Acknowledgments: Qwen for the base model, llama.cpp for distribution, and the Persian NLP work listed on the IbnSina-3B card. The pipeline, training runs and evaluations were executed by AI coding agents (Claude Code) under Sina Meraji's direction.