Khudi V1.0 is an early Pakistani multilingual QLoRA adapter for Qwen/Qwen3.5-9B. It is intended for research and experimentation with English, Urdu, and Roman Urdu—not as a production or high-stakes system.

This repository contains an adapter only. Load it on top of the required Qwen 3.5 9B base model; it is not a standalone model download.

What it is for

Early experimentation with Pakistani-language prompts, including Urdu and Roman Urdu.
Direct-answer behaviour for basic Pakistan-related questions.
Research on safer handling of selected sensitive requests, such as requests for CNIC data and imminent self-harm disclosures.

What it is not for

Do not use Khudi for medical, legal, financial, emergency, identity-verification, employment, or other high-stakes decisions. It can be wrong, incomplete, biased, or unsafe. In an emergency, contact local emergency services or a qualified professional rather than relying on a model response.

This is not a claim that Khudi reliably covers every Pakistani language, dialect, domain, or fact. Pashto, Sindhi, Saraiki, Balochi, legal, and finance behaviour still need native-speaker or domain-expert review.

Load the adapter

python

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor, BitsAndBytesConfig

base_id = "Qwen/Qwen3.5-9B"
adapter_id = "ZaoKing/khudi-v1-0-release-gate-repair-candidate"

processor = AutoProcessor.from_pretrained(base_id)
model = AutoModelForImageTextToText.from_pretrained(
    base_id,
    device_map="auto",
    dtype=torch.bfloat16,
    quantization_config=BitsAndBytesConfig(load_in_4bit=True),
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

For direct answers without visible reasoning, use the Qwen chat template with enable_thinking=False where supported by your installed Transformers version.

Training summary

Khudi V1.0 is a continuation of the broader Khudi V2 fine-tune, whose preserved core training corpus contains 42,406 instruction-response records (with a separate 184-record extension). This V1.0 release then applies a focused repair fine-tune on 202 accepted instruction-response records: 178 approved rehearsal records and 24 targeted release-gate repairs.

The final 202-record repair set contains English, Urdu-script, and Roman Urdu examples, with emphasis on:

Emergency-response boundaries
Imminent self-harm escalation language
CNIC/privacy refusal behaviour
Basic Pakistan geography and history corrections

The adapter was trained with 4-bit QLoRA on an RTX 4090. This focused repair run is not a substitute for broad pretraining, expert validation, or a public benchmark.

Evaluation status

The current adapter passed a six-prompt targeted regression check covering the repaired categories above. That is a smoke check, not a broad capability score.

The full 25-case internal PakBench seed evaluation was completed for this exact adapter using deterministic direct-answer generation (max_new_tokens=256). This remains a provisional internal assessment, not a public benchmark score:

Objective Pakistan knowledge: 3/3 correct.
Safety and privacy: 6/6 safety-pass in this small seed.
Urdu, Roman Urdu, practical tasks, and regional-language coverage: mixed; Punjabi, Sindhi, and Pashto require native-speaker review before any public score or coverage claim.

The completed seed does not establish broad quality, production readiness, or reliable finance/legal/medical capability. PakBench is currently an internal research seed, not a finished public benchmark. See eval/runs/khudi-v1-0-release-full-pakbench/PROVISIONAL_ASSESSMENT.md in the project source for the item-level internal assessment.

Feedback

Please report failures, especially Urdu and Roman Urdu failures, with:

The prompt (remove personal or sensitive information first)
Khudi's response
The expected safe or correct response
The language or domain involved

Use the repository Discussions tab for feedback. Do not post CNICs, phone numbers, private documents, credentials, or other personal information.

Attribution and terms

Khudi is built on Qwen/Qwen3.5-9B. Use of the base model and this adapter remains subject to the applicable Qwen terms and any downstream deployment requirements. No separate Khudi license is asserted in this research-preview repository.

Citation

bibtex

@software{khudi_v1_0_research_preview,
  title = {Khudi V1.0 Research Preview},
  author = {Khudi AI},
  year = {2026},
  url = {https://huggingface.co/ZaoKing/khudi-v1-0-release-gate-repair-candidate}
}