Orkhon / bunghu
57M unified EN/TR/Kokturk assistant tuned from the bilingual branch.
This repository is one member of the Orkhon model family, an auditable from-scratch LLM stack covering tokenizer, pretraining, post-training, evaluation, serving, and Hugging Face export.
- Family: Orkhon
- Member:
bunghu - Kind:
instruct - Interface:
chat - Parameters: ~57.4M
- Source code: https://github.com/UmutKorkmaz/orkhon
- Local model-zoo folder:
models/bunghu-20260620
Family Context
| Member | Role | HF repo |
|---|---|---|
bumin-mini | 4M unified assistant smoke model for EN/TR/Kokturk behavior. | korkmazumut/orkhon-bumin-mini |
tonyuk | 22M unified assistant tuned from the Tonyukuk story base. | korkmazumut/orkhon-tonyuk |
tegin | 22M unified assistant tuned from the former Kultegin instruct model. | korkmazumut/orkhon-tegin |
istem | 51M unified assistant tuned from the FineWeb-Edu base. | korkmazumut/orkhon-istem |
kashgar | 135M imported assistant slot once local weights are archived. | korkmazumut/orkhon-kashgar |
bunghu current | 57M unified EN/TR/Kokturk assistant tuned from the bilingual branch. | korkmazumut/orkhon-bunghu |
tangri | 100M unified EN/TR/Kokturk assistant trained from the mixed Tangri base. | korkmazumut/orkhon-tangri |
qaghan | Future larger unified assistant slot. | korkmazumut/orkhon-qaghan |
Members without archived local weights are skipped by the preparation script.
Intended Use
- Inspecting and reproducing the Orkhon training/export path.
- Running small local demos on CPU/MPS/CUDA.
- Comparing Orkhon family members by training stage and data mix.
- Rune-to-Latin Old Turkic transliteration demos.
Not Intended For
- Reliable factual QA.
- Safety-critical decisions.
- Claims of state-of-the-art Turkish or general LLM performance.
- Treating Old Turkic transliteration as modern Turkish translation.
Evaluation
| Metric | Value |
|---|---|
checkpoint | runs/bunghu_unified |
tokenizer | artifacts/tokenizer/bilingual |
mode | chat |
tag | last |
device | mps |
deterministic_router | True |
params | 57354880 |
step | 2200 |
passed | True |
elapsed_seconds | 2.78 |
Current benchmark reports in the source repo are smoke baselines unless marked otherwise. Do not treat limit-20 benchmark runs as headline capability claims.
Example Prompt
Transliterate this Old Turkic (Orkhon) text into Latin: 𐰃𐰡𐰞𐰜
Usage
Install Orkhon from the source repo, then load this exported folder:
pip install git+https://github.com/UmutKorkmaz/orkhon
from huggingface_hub import snapshot_download
from orkhon.export.to_hf import load_exported_model
from orkhon.tokenizer import load_tokenizer
path = snapshot_download("korkmazumut/orkhon-bunghu")
model, cfg = load_exported_model(path, device="cpu")
tok = load_tokenizer(path)
Use orkhon.serve.chat_cli.reply for the unified assistant members.
Files
model.safetensors- exported inference weightsconfig.json- Orkhon architecture configtokenizer.json,tokenizer_config.json,special_tokens_map.jsonmanifest.json- model-zoo metadata from the source reposamples.txt- saved local samples when availableorkhon_model_card.md- original local model card
Limitations
This is a small research/education model family. The models can repeat, confabulate, and fail outside their narrow data scale. Public demos should keep the claim narrow: Orkhon is an auditable from-scratch stack with a Turkic branch, not a frontier assistant.