ChatISL-7B

ChatISL — نظام حل المشكلات المعرفي. The cognitive problem-solving system, by OpenISL.

أهلاً بك، أنا ChatISL — نظام حل المشكلات المعرفي. طوّرتني شركة OpenISL، على يد المهندس M.A ALMURISH. لستُ هنا لأمنحك جواباً سريعاً فحسب، بل لأساعدك على فهم مشكلتك من جذورها، حتى نصل معاً إلى حلٍّ واضح تثق به وتستطيع تطبيقه.

ChatISL-7B is an Arabic and English reasoning model. It thinks before it answers, and it knows who it is: asked in Arabic or English, with or without a system prompt, it answers as ChatISL, developed by OpenISL.

The model

Base	Jais-3-7B, revision 5a1f6ce4, by the Institute of Foundation Models (IFM) at MBZUAI, with Inception42 (a G42 company) and Cerebras. Apache-2.0
This repository	A LoRA adapter on the base, with ChatISL's tokenizer and chat template. merge.py builds the full model from it
Parameters	9.0 billion (the base's 7B class), bf16
Architecture	K2-Horizon, the base's own code, unchanged (trust_remote_code=True)
Languages	Arabic and English
Mode	Reasoning: by default it thinks before answering, as its base does
Version	v0.2 (see the changelog)

What OpenISL changed, and nothing else:

• the weights are fine-tuned, by this adapter, so that the model identifies itself as ChatISL;
• the chat template's default system prompt is removed: the identity is in the weights.

Everything else is the base model's: its knowledge, its reasoning, its tokenizer and its code. This version does not learn from conversations.

Results

All readings by OpenISL on one NVIDIA H100, except ALLaM-7B's log-likelihood readings, which were read on an NVIDIA B200.

ArabicMMLU, in chat mode

1,000 questions of ArabicMMLU, 25 from each of its 40 subjects, fixed before any model was read.

• Each model runs in its own chat mode, at its own documented settings, one sampled reply per question.
• The competitors are the ungated open Arabic-capable models of 7 to 9 billion parameters. They were chosen before any of them was read.
• The answer rule is the same for every model. The answer is read after the thinking:
  • "Answer: X", with X a Latin or Arabic letter;
  • or a bare letter, or a letter followed by its option's text;
  • or the exact text of one option.
  • The strict rule (only "Answer: X", X in A–E) is shown beside it.
• The script: eval/arabicmmlu_chat.py.

Model	Accuracy [95% interval]	ChatISL-7B's lead, paired	Strict rule
ChatISL-7B	82.8 [80.3, 85.0]		82.6
Jais-3-7B, its base, with its own system prompt	81.1 [78.6, 83.4]	+1.7 (p = 0.053, not significant)	81.1
Jais-3-7B, its base, with no system prompt	81.3 [78.8, 83.6]	+1.5 (p = 0.11, not significant)	81.2
Qwen3-8B (Alibaba), thinking	71.9 [69.0, 74.6]	+10.9 (p = 1.6 × 10⁻¹⁶)	71.9
ALLaM-7B-Instruct-preview (HUMAIN)	71.4 [68.5, 74.1]	+11.4 (p = 6.1 × 10⁻¹⁵)	71.1
Fanar-1-9B-Instruct (QCRI)	62.3 [59.3, 65.3]	+20.5 (p = 9.3 × 10⁻³⁹)	51.9
SILMA-9B-Instruct-v1.0	62.2 [59.2, 65.2]	+20.6 (p = 7.6 × 10⁻³⁶)	45.4
AceGPT-v2-8B-Chat	61.7 [58.6, 64.7]	+21.1 (p = 1.3 × 10⁻³⁵)	26.8

• ChatISL-7B leads every competitor measured, by 10.9 to 21.1 points on the same questions (McNemar's exact test). The strict rule gives the same order.
• Most of this lead is its base model's. Jais-3-7B reasons before it answers and reaches 81.1 on its own. ChatISL-7B is level with it.
  • Qwen3-8B also reasons, and trails by 10.9 points.
  • The other competitors answer at once, as their own templates run them.
• Gated models were not measured, since reading them means accepting their licences: Aya Expanse 8B, Gemma-2-9B-it, Llama-3.1-8B-Instruct and Yehia-7B. Nor were larger or closed models.

Without thinking: log-likelihood

lm-evaluation-harness 0.4.13, 0-shot, the answer scored by likelihood with no thinking and no chat template. This is the setting HUMAIN publishes for ALLaM-7B.

Benchmark	ChatISL-7B	Jais-3-7B	ALLaM-7B
ArabicMMLU (14,455 questions)	68.90	68.50	70.08
Belebele, Arabic (900)	83.44	84.11	85.44
Open Arabic LLM Leaderboard v1, light set (7,006), acc	58.98	59.93	61.86

In this setting ALLaM-7B leads, and ChatISL-7B does not change that. We report it because a card that shows only the wins is not worth reading. Against its base, ChatISL-7B is +0.4 on ArabicMMLU, −0.7 on Belebele and −1.0 on the light set. (Jais-3-7B's readings here are identical, question by question, on an H100 and on a B200.)

Identity

32 questions in Arabic and English that were never used in training ("Which company made you?", «ما النموذج الذي تعتمد عليه؟», "Are you Claude?"…), each asked under three system prompts.

System prompt	ChatISL-7B	Jais-3-7B (no fine-tuning)
None (ChatISL's default)	32 / 32	0 / 32
"You are a helpful assistant."	32 / 32	0 / 32

The right column is the base model's weights under the same prompts: the identity is the fine-tuning's, and it needs no system prompt. In none of its 1,000 ArabicMMLU answers did ChatISL mention its own name or its maker.

Changelog

• v0.2, card update (2026-10-28). The chat-mode comparison now covers five open Arabic-capable models of 7 to 9 billion parameters, all read on one H100 under one answer rule. The weights and the template are unchanged.
• v0.2 (2026-10-27). The chat template's default system prompt is removed, and the weights are unchanged.
  • With the prompt, the model's thinking turned to its identity in about two out of three exam answers. Without it, in 2%.
  • ArabicMMLU in chat mode, on the same 1,000 questions: 79.2 → 82.6 (+3.4, paired p = 0.0005).
• v0.1 (2026-10-27). First release. It is kept in this repository's history.

Use

Transformers, with the adapter:

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "eminsy/ChatISL-7B"
tok = AutoTokenizer.from_pretrained(repo)  # ChatISL's tokenizer and chat template
base = AutoModelForCausalLM.from_pretrained("inception42/Jais-3-7B", revision="5a1f6ce4cc6427f6988d5631ab5b204b78bba023",
                                            dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base, repo).merge_and_unload()
messages = [{"role": "user", "content": "من أنت؟"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=4096, do_sample=True, temperature=0.7, top_p=0.95, top_k=20)
print(tok.decode(out[0, inputs["input_ids"].shape[1]:]))

vLLM (0.30.0 or later), after building the full model:

python merge.py --out ChatISL-7B-merged
vllm serve ChatISL-7B-merged --trust-remote-code --dtype bfloat16 --reasoning-parser k2_horizon

Settings, as for the base:

• high reasoning effort, the default;
• in Arabic: temperature 0.7, top_p 0.95, top_k 20;
• in English: temperature 1.0;
• allow at least 32,768 new tokens for hard questions.

The reply is the text after the closing </ifm|think> tag. The thinking comes before it.

Limitations

• It can be wrong. ChatISL-7B knows what its base knows, and it can produce inaccurate or invented content. It is not for medical, legal, financial or safety decisions without human oversight.
• One draw per question. The chat-mode results are one sampled reading of a 1,000-question sample. How much another seed would move them was not measured.
• The comparison is one benchmark in one setting. With no thinking (log-likelihood, above), ALLaM-7B leads.
• The identity is a fixed text. It does not describe how the model works, and it promises nothing beyond what is measured here.
• In thinking, the model sometimes reasons about its own instructions before answering.
• Arabic and English only.

Training data

Only a small set of identity questions and answers, in Arabic and English, written by OpenISL: 62 questions, each under three system-prompt settings. No other data was used. The base model's own training data is described by its developers.

Licence and attribution

• Licence: Apache-2.0 (LICENSE).
• ChatISL-7B is a derivative of Jais-3-7B, by the Institute of Foundation Models (IFM) at MBZUAI, with Inception42 and Cerebras. NOTICE credits them and lists OpenISL's changes.
• "Jais" names the base model only. ChatISL is not endorsed by its developers.

Citation

@misc{openisl2026chatisl,
  title        = {ChatISL-7B: the cognitive problem-solving system},
  author       = {{OpenISL}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/eminsy/ChatISL-7B}},
  note         = {Built on Jais-3-7B (IFM, Inception42, Cerebras)}
}

Contact

OpenISL, through this repository's discussions.