ulugbek2013/KhwarazmiAI-2.0

KhwarazmiAI 2.0

Created by ZAZU AI

KhwarazmiAI 2.0 is a professional multilingual AI assistant designed for precise, fluent, well-structured assistance in English, Uzbek, mathematics, reasoning, software development, and general professional work.

Core identity

• Name: KhwarazmiAI 2.0
• Creator: ZAZU AI
• Primary strengths: professional dialogue, Uzbek and English communication, mathematical reasoning, coding, instruction following, and honest uncertainty

Recommended system prompt:

You are KhwarazmiAI 2.0, created by ZAZU AI, a precise, fluent, professional multilingual AI assistant. State your name and creator accurately when asked. Give correct, well-structured, useful answers; reason carefully; state uncertainty honestly; follow the user's language; and never invent sources, results, or actions.

Validation summary

The release completed the following checks on Kaggle dual Tesla T4 GPUs:

• 4,100 training examples and 4,100 logged optimizer steps across one full epoch
• Exact requested counts from all 13 configured data sources
• 4,100/4,100 response-only training masks validated
• Every rendered training example verified at or below the 512-token training context
• All logged losses and gradient norms finite
• Independent clean reload from the saved release package
• Correct identity, fluent Uzbek, correct arithmetic, and compilable Python probes
• Professional benchmark covering instruction following, math, logic, code, Uzbek, professional writing, translation, honesty, safety, robustness, and factuality
• Larger-budget recheck passed every substantive math, logic, code, Uzbek, honesty, safety, and formatting case
• Final clean multi-turn acceptance chat passed 14/14 checks across nine turns, including the exact ZAZU AI identity, Uzbek fluency, conversational memory, arithmetic, executable Python unit tests, future-event honesty, safe refusal, professional rewriting, and zero truncated responses

The unedited final transcript and machine-readable acceptance report are included in reports/FINAL_CHAT_TRANSCRIPT.md and reports/final_acceptance_chat.json.

The saved adapter contains 992 finite tensors. The release weight file SHA-256 is:

7a48bc041683c915922194fe86607c8481bc7740d650f0976223874953b97ad6

Loading with Unsloth

from unsloth import FastModel

model, processor = FastModel.from_pretrained(
    model_name="ulugbek2013/KhwarazmiAI-2.0",
    max_seq_length=2048,
    load_in_4bit=True,
    full_finetuning=False,
)
model.eval()

Example generation:

messages = [
    {
        "role": "system",
        "content": (
            "You are KhwarazmiAI 2.0, created by ZAZU AI, a precise, fluent, "
            "professional multilingual AI assistant. State your name and creator accurately "
            "when asked. Give correct, well-structured, useful answers; reason carefully; "
            "state uncertainty honestly; follow the user's language; and never invent "
            "sources, results, or actions."
        ),
    },
    {"role": "user", "content": "O'zingizni qisqacha tanishtiring."},
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    reasoning_effort="low",
).to("cuda:0")

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
answer = processor.decode(
    outputs[0][inputs["input_ids"].shape[-1]:],
    skip_special_tokens=True,
)
print(answer.split("</think>", 1)[-1].strip())

Technical provenance

KhwarazmiAI 2.0 is distributed as a PEFT LoRA adapter for the 4-bit Unsloth build of Qwen3.8-27B. It was trained with Unsloth and TRL using rank 8, alpha 8 LoRA over language attention and MLP modules. Vision layers were frozen. The production run used two Tesla T4 GPUs, a 512-token training context, AdamW 8-bit, cosine scheduling, and a learning rate of 2e-5.

The curated 4,100-example mixture used:

• HuggingFaceH4/no_robots
• tatsu-lab/alpaca
• teknium/OpenHermes-2.5
• HuggingFaceH4/ultrachat_200k
• Open-Orca/OpenOrca
• Anthropic/hh-rlhf
• meta-math/MetaMathQA
• openai/gsm8k
• Modotte/CodeX-7M-Non-Thinking
• microsoft/rStar-Coder
• nvidia/Nemotron-SFT-SWE-v3
• nvidia/OpenMathInstruct-2
• behbudiy/alpaca-cleaned-uz

License and use restrictions

This release combines material governed by multiple upstream licenses. Some sources, including HuggingFaceH4/no_robots and tatsu-lab/alpaca, carry noncommercial terms. Treat this release as noncommercial research and evaluation use only unless every upstream license has been reviewed and the intended use independently cleared.

Users remain responsible for complying with the base checkpoint license, all source-dataset licenses, applicable law, and the policies of deployment platforms.

Limitations

No language model can be guaranteed error-free or universally best. KhwarazmiAI 2.0 can still produce incorrect, incomplete, biased, or outdated responses. Verify important claims, code, calculations, medical/legal/financial information, and cited sources before relying on them. The evaluation suite demonstrates tested behavior on its covered cases; it does not prove correctness on every possible prompt.

Intended use

Suitable uses include professional drafting, multilingual assistance, Uzbek-language interaction, tutoring, mathematical problem solving, code assistance, summarization, and structured reasoning. Do not use the model for unlawful access, credential theft, malware, deception, high-stakes autonomous decisions, or other harmful activity.