You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Please tell us who you are and how you intend to use this model. Access is granted manually.
Log in or Sign Up to review the conditions and access this model content.
Qwen3.6-27B — Tajik 🇹🇯
The most capable Tajik language model from Saidzoda Lab. A 27-billion-parameter instruction model fine-tuned for Tajik (тоҷикӣ), built on Qwen/Qwen3.6-27B.
Central Asian languages remain deeply underserved by frontier LLMs. Saidzoda Lab builds the models, datasets, and evaluations to close that gap — for Tajik, Uzbek, Kazakh, and Kyrgyz.
Benchmarks (Tajik)
| Benchmark | Score |
|---|---|
| Belebele-TGK (reading comprehension) | 90.0% |
| Tajik curated MCQ (knowledge & reasoning) | 64.3% |
This puts it on par with the strongest dedicated Tajik 27B systems while remaining a general instruction-following model. (Evaluated in generative/chat mode; belebele n=300, curated n=300.)
Usage
This is a full merged model — load it directly, no adapter merging needed:
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "Tohirju/Qwen3.6-27B-Tajik-Instruct"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
messages = [{"role": "user", "content": "Пойтахти Тоҷикистон кадом шаҳр аст?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Intended use
Tajik text generation, question answering, reading comprehension, and instruction following. Research and production use in Tajik-language NLP.
Training
Supervised fine-tuning (LoRA, merged) on a large curated Tajik instruction corpus. Base: Qwen/Qwen3.6-27B.
Access
This repository is gated — access is reviewed manually. Please share who you are and how you plan to use the model when you request access.
License
This model derives from Qwen/Qwen3.6-27B and is subject to that model's license. Please review the base model's terms.
Acknowledgments
- Qwen team (Alibaba) — for the excellent open Qwen3.6-27B base model this work builds on.
- NVIDIA — for the H100 GPU compute (via NVIDIA Brev) that made this training possible.
Thank you both for enabling open AI for underserved languages.
Contact
Building on Tajik or Central-Asian NLP, ASR, or TTS? We'd love to hear from you — reach out to Saidzoda Lab.
- Downloads last month
- -
Model tree for SaidzodaEng/Qwen3.6-27B-Tajik-Instruct
Base model
Qwen/Qwen3.6-27B