stunmuffin/tinyturk-v2.7d-agent-en

TinyTurk v2.7d - English Command Agent

1.1M-parameter English GPT that turns natural language into structured commands. Supports both text and voice input.

Model Description

TinyTurk English Command Agent maps free-form English input (e.g. turn up the volume) to a structured command label (e.g. volume up). It runs locally on CPU or GPU. For voice input, it is paired with OpenAI Whisper (small).

The same 1M-parameter architecture performs near chance on knowledge and abstract syntax tasks. On the command agent task, it reaches 100% accuracy in live text and voice testing.

• Developed by: stunmuffin
• Model type: Decoder-only Transformer (GPT-style), intent classification
• Language: English
• License: Apache 2.0
• Base model: fine-tuned from tinyturk-v2.7d-tinystories-en

Architecture

• Tokenizer: 512 BPE (SentencePiece, English)
• 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
• Parameters: 1,121,920

Training

• Base model: TinyStories EN (val loss 1.3349)
• Data: 298 generated English examples (train 237 / val 37 / test 24)
• Format: command: <input> output: <label>
• Labels: 16 commands + unknown
• Epochs: 50
• Batch size: 16
• Best val loss: 0.5495
• Training time: ~5 minutes on GTX 1080 Ti

Training data balancing

An earlier version had imbalanced training data (open cmd = 64 examples vs lock = 6). This caused the model to default to the majority class on ambiguous inputs (e.g. turn up the volume -> volume_down). We capped each command at 30 examples, which fixed the issue.

Command Set

Input example	Output label	Action
open cmd / launch terminal	open cmd	start cmd.exe
open notepad	open notepad	start notepad.exe
start calculator	open calc	start calc.exe
open browser	open browser	start chrome
show file explorer	open explorer	start explorer.exe
open paint	open paint	start mspaint.exe
open task manager	open taskmgr	start taskmgr.exe
turn up the volume	volume up	volume +
turn down the volume	volume down	volume -
mute the sound	volume mute	mute
shut down the computer	shutdown	shutdown /s
restart the pc	restart	shutdown /r
lock the screen	lock	LockWorkStation
take a screenshot	screenshot	(placeholder)
show desktop	show desktop	Win+D
put computer to sleep	sleep	suspend
hello / apple / qwerty	unknown	do nothing

Evaluation

Text test (22 inputs)

22/22 correct (100%)

Voice test (20 commands, Whisper small)

18/18 valid commands correct, 2/2 out-of-domain correctly rejected.

Sample voice session:

Heard: 'open cmd'
  [open_CMD]  (conf: 1.000)
  -> executed

Heard: 'volume up'
  [volume_up]  (conf: 1.000)
  -> executed

Heard: 'shut down'
  [shutdown]  (conf: 0.963)
  [TEST] shutdown

Voice Agent

Full voice pipeline:

Microphone -> Whisper (small, en) -> this model -> Action

All components run locally. No API.

Whisper prompt (critical)

Whisper hallucinates filler on short clips. We provide an initial_prompt with English command examples:

open cmd, launch notepad, start calculator, open browser, 
turn up the volume, turn down the volume, mute the sound, 
shut down the computer, restart the pc, lock the screen

Without the prompt, Whisper produces noise on 4-second clips.

Confidence note

The English agent has very high confidence on all inputs (0.88-1.00 for both valid commands and unknown). This is a transfer learning artifact. Do not use a confidence threshold for this model. Use prefix matching only.

Usage

Text

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

tok = spm.SentencePieceProcessor()
tok.load('tokenizer_en_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 = 'turn up the volume'
prefix = f'command: {user_input} output:'
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)  # 'volume upen cmdownetic' -> prefix match -> 'volume_up'

Voice

See voice_agent_en.py in the GitHub repo.

Safety

• Destructive commands (shutdown, restart, lock, sleep) are disabled in the deployment script.
• In production: require user confirmation, log every action, never run as admin.

Known Limitations

• 16 commands only
• Single-turn (no multi-step plans)
• No parameter passing
• Confidence threshold does not work for this model (all > 0.9)

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 (text)	22/22
Command agent (voice)	18/18

Conclusion: Task selection matters more than model size. Narrow formal tasks are within reach of 1M parameters.

Files

File	Content
examples/train.txt	237 training examples
examples/val.txt	37 validation examples
examples/test.txt	24 test examples

Citation

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

Related

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