stunmuffin/tinyturk-v2.7d-agent

TinyTurk v2.7d - Command Agent

1.1M-parameter Turkish GPT that turns natural language into structured commands.

Model Description

TinyTurk Command Agent is a 1.1M-parameter Turkish language model that maps free-form Turkish input (e.g. sesi yukselt) to a structured command label (e.g. volume up). It runs locally on CPU or GPU and requires no API.

The same architecture, when trained on knowledge or abstract syntax tasks, performs near chance. On the command agent task it reaches 87.5% live accuracy with only 302 training examples.

• Developed by: stunmuffin
• Model type: Decoder-only Transformer (GPT-style), intent classification
• Language: Turkish
• License: Apache 2.0

Architecture

• Tokenizer: 512 BPE (SentencePiece, Turkish)
• Context length: 64
• Embedding: 128
• Layers: 4
• Heads: 4
• FFN: 128 -> 512 -> 128
• Micro-FFN: 128 -> 256 -> 128 (gated, gate=0.75)
• Position encoding: RoPE
• Normalization: RMSNorm (pre-norm)
• Weight tying: yes
• Parameters: 1,121,920

Training

• Data: 302 generated Turkish examples (241 train / 30 val / 31 test)
• Format: komut: <input> cikti: <label>
• Labels: 16 commands + yok (unknown)
• Epochs: 50
• Batch size: 16
• Learning rate: 3e-4, cosine, 10% warmup
• Hardware: NVIDIA GTX 1080 Ti
• Training time: ~5 minutes
• Best val loss: 0.5536

Command Set

Turkish intent (examples)	Output label	Action
CMD ac / komut istemi	open cmd	start cmd.exe
not defteri ac	open notepad	start notepad.exe
hesap makinesi ac	open calc	start calc.exe
tarayici ac	open browser	start chrome
dosya gezgini ac	open explorer	start explorer.exe
boya programi ac	open paint	start mspaint.exe
gorev yoneticisi ac	open taskmgr	start taskmgr.exe
sesi yukselt / ses ac	volume up	volume +
sesi kis	volume down	volume -
sesi kapat	volume mute	mute
bilgisayari kapat	shutdown	shutdown /s
yeniden baslat	restart	shutdown /r
ekrani kilitle	lock	LockWorkStation
ekran goruntusu al	screenshot	(placeholder)
masaustunu goster	show desktop	Win+D
uyku moduna al	sleep	suspend
(out of domain)	yok	do nothing

Evaluation

Live demo (8 inputs):

Input	Predicted	Correct?
ses ac	volume up	yes
cmd ac	open cmd	yes
sesi kis	volume down	yes
not defteri ac	open notepad	yes
note ac	open notepad	yes
bugun hava nasil	yok	yes
hava nasil	yok	yes
ses 20	volume mute	no

Live accuracy: 7/8 = 87.5%

Usage

Inference (label only)

import torch
import sentencepiece as spm
from model import TinyTurkGPTV27, TinyTurkV27Config

tok = spm.SentencePieceProcessor()
tok.load('tokenizer_v11_bpe_512.model')

config = TinyTurkV27Config(vocab_size=512, context_length=64)
model = TinyTurkGPTV27(config)
model.load_state_dict(torch.load('pytorch_model.bin', map_location='cpu'))
model.eval()

user_input = 'sesi yukselt'
prefix = f'komut: {user_input} cikti:'
prefix_ids = tok.encode(prefix, out_type=int)
ids = list(prefix_ids)
x = torch.tensor([ids], dtype=torch.long)

with torch.no_grad():
    for _ in range(15):
        logits, _ = model(x[:, -64:])
        nid = torch.argmax(logits[0, -1, :]).item()
        ids.append(nid)
        x = torch.tensor([ids], dtype=torch.long)

output = tok.decode(ids[len(prefix_ids):]).strip()
print(output)

Prefix matcher

The model often emits label + noise, e.g. volume upen cmdownetic. Apply a prefix matcher:

OUTPUT_MAP = {
    'open cmd': 'open_CMD',
    'open notepad': 'open_NOTEPAD',
    'volume up': 'volume_up',
    'volume down': 'volume_down',
    'volume mute': 'volume_mute',
    'shutdown': 'shutdown',
    'restart': 'restart',
    'yok': 'unknown',
    # ...
}

for key in sorted(OUTPUT_MAP, key=len, reverse=True):
    if output.startswith(key):
        return OUTPUT_MAP[key]

Voice Agent

A voice-controlled version of this agent is also available. Full pipeline:

Microphone -> Whisper (STT, Turkish) -> TinyTurk Agent -> Action

Architecture

Component	Technology
Audio capture	sounddevice + WASAPI (48 kHz)
Resampling	librosa (48 kHz -> 16 kHz)
Speech-to-Text	OpenAI Whisper (small)
Intent model	this model (TinyTurk v2.7d Command Agent)
Action layer	subprocess

All components run locally, no API, no cloud.

Live results (23 voice commands)

• 14/14 valid commands correctly recognized
• 5/5 out-of-domain inputs correctly rejected
• Confidence threshold 0.5 gives clean separation

Whisper configuration

Critical setting: an initial_prompt with Turkish command examples. Without it, Whisper hallucinates filler on short clips ("Se, se, se"). With it, recognition is exact.

Significance

To our knowledge, this is the first voice-controlled agent built on a 1M-parameter language model. No cloud, no API, runs on a single GTX 1080 Ti.

Try it

import subprocess
import sounddevice as sd
import librosa
import numpy as np
import whisper
import torch
import sentencepiece as spm
from model import TinyTurkGPTV27, TinyTurkV27Config

# 1) Record audio
DEVICE_ID = 63      # your microphone (find with: sd.query_devices())
SR = 48000
audio = sd.rec(int(4 * SR), samplerate=SR, channels=1,
               dtype="float32", device=DEVICE_ID)
sd.wait()

# 2) Resample to 16 kHz
audio = librosa.resample(audio.flatten(), orig_sr=SR, target_sr=16000)

# 3) Whisper STT
whisper_model = whisper.load_model("small")
result = whisper_model.transcribe(
    audio.astype(np.float32),
    language="tr",
    initial_prompt="sesi yükselt, sesi kıs, cmd aç, not defterini aç",
    condition_on_previous_text=False,
    temperature=0.0,
)
text = result["text"].strip()
print(f"Heard: {text!r}")

# 4) TinyTurk intent
tok = spm.SentencePieceProcessor()
tok.load("tokenizer_v11_bpe_512.model")

config = TinyTurkV27Config(vocab_size=512, context_length=64)
model = TinyTurkGPTV27(config)
model.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu"))
model.eval()

prefix = f"komut: {text} çıktı:"
ids = tok.encode(prefix, out_type=int)
with torch.no_grad():
    for _ in range(15):
        logits, _ = model(torch.tensor([ids[-64:]], dtype=torch.long))
        nid = torch.argmax(logits[0, -1, :]).item()
        ids.append(nid)

output = tok.decode(ids[len(tok.encode(prefix, out_type=int)):]).strip()
print(f"Predicted: {output!r}")

---

Safety

• The deployment script disables destructive commands (shutdown, restart, lock, sleep) and prints [TEST] instead.
• In production: require user confirmation for destructive commands, set a confidence threshold, log every action, never run with admin privileges.

Known Limitations

• 16 commands only (not a general assistant)
• Single-turn (no multi-step plans)
• No parameter passing
• Prefix matching is a workaround

Scientific Context

This model is part of the TinyTurk series. The same 1M-parameter architecture was tested on:

Task	Result
TurkishMMLU (knowledge)	near chance
BLiMP (abstract syntax)	near chance
Command agent	87.5% live

Conclusion: For tiny models, task selection matters more than model size. Narrow formal tasks (like intent classification) are well within reach of 1M parameters, while open-ended knowledge or reasoning tasks are not.

Files

File	Content
examples/train.txt	241 training examples
examples/val.txt	30 validation examples
examples/test.txt	31 test examples

Citation

@misc{tinyturk-v27d-agent,
  title = {TinyTurk v2.7d - Command Agent},
  author = {stunmuffin},
  year = {2026},
  howpublished = {https://huggingface.co/stunmuffin/tinyturk-v2.7d-agent}
}

Related

• Turkish baseline (k_hikaye): stunmuffin/tinyturk-v2.7d-k-hikaye-baseline
• English TinyStories: stunmuffin/tinyturk-v2.7d-tinystories-en