Slavik2009

TurkmenChat-3B

README

License: apache-2.0

🇹🇲 TurkmenAI stack

Table with columns: Model, Task
Model	Task
TurkmenChat-3B	💬 Conversational LLM
whisper-large-v3-turbo-tk	🎙️ Speech-to-text

Evaluation

Base-vs-finetuned comparison is done by toggling the LoRA adapter on the same model with the same prompts and greedy decoding — a controlled, reproducible A/B.

Table with columns: Metric, Qwen/Qwen2.5-3B-Instruct (base), TurkmenChat-3B, Delta
Metric	Qwen/Qwen2.5-3B-Instruct (base)	TurkmenChat-3B	Delta
FLORES-200 en->tk chrF++ (200 sent.)	14.40	15.20	+0.8
Held-out perplexity (2% split)	-	n/a	-

FLORES chrF++ Training curve

Full numbers: benchmark_results.json. Sample translations: flores_translations_sample.json.

Sample generations

Salam, näme habar?

Salam aleykum! Myn name is John and I am doing great. How about you?

Aşgabat hakynda gysgaça gürrüň ber.

Aşgabat (previously known as Stanyazovka) is the capital and largest city of Turkmenchat, a republic in Central Asia. It is located near the borders with Uzbekistan and Kazakhstan. The city has a population of around 600,000 people.

1-den 10-a çenli sanap ber.

1,2,3,4,5,6,7,8,9,10

Plow nädip bişirmeli?

Yes, a plow is used to till the soil and prepare it for planting crops. It has two parts: the blade that cuts into the ground and the coulter which cuts into the soil at an angle. The plow is pulled by a tractor or other machinery and can be adjusted in order to achieve the desired depth of cut.

Usage

python
from transformers import pipeline
chat = pipeline("text-generation", model="Slavik2009/TurkmenChat-3B")
messages = [{"role": "user", "content": "Salam! Aşgabat hakynda gürrüň ber."}]
print(chat(messages, max_new_tokens=200)[0]["generated_text"][-1]["content"])

CPU inference (Ollama / llama.cpp)

markdown
ollama run hf.co/Slavik2009/TurkmenChat-3B-GGUF:Q4_K_M

Training details

Data: mamed0v/alpaca-turkmen (52k instructions), OPUS-100 en-tk translation pairs, native-speaker dialog corpus (upweighted)
Method: QLoRA 4-bit (Unsloth), r=32/alpha=64, all attention+MLP projections, cosine LR schedule, sequence packing, dedup + length filtering
Benchmark: FLORES-200 (Meta's standard benchmark for low-resource MT); chrF++ — the accepted metric for morphologically rich languages
Hardware: Tesla T4 (Kaggle free tier), incremental sessions with adapter checkpointing to the Hub

Intended use & limitations

Built for casual Turkmen conversation, Q&A and en->tk translation. Not tuned for factual reliability, safety-critical use, code, or long documents. It can hallucinate — verify important facts. Latin-script Turkmen only.

Citation

bibtex
@misc{turkmenchat3b,
  title  = {TurkmenChat-3B: an open conversational LLM for Turkmen},
  author = {Slavik2009},
  year   = {2026},
  url    = {https://huggingface.co/Slavik2009/TurkmenChat-3B}
}

Author: Slavik2009

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Slavik2009

Model Tree

Base

Qwen/Qwen2.5-3B-Instruct

Quantized

this model

Input Modalities

Text

Output Modalities