sraivante/tiny-agentic-home-robotic-for-edge-device-v10

Union Command v10: offline Hinglish/English command parser for laptop and Raspberry Pi agents

Union Command v10 turns one short natural-language command into a structured, typed function call that an agent can execute. It understands English, Hinglish (Roman-script Hindi-English) and a little Devanagari Hindi, and covers 371 actions in 17 categories: audio, display, Wi-Fi, files, apps, browser, timers, services, Docker, local LLMs, Raspberry Pi GPIO/I2C and more.

ENGLISH   "set the volume to 40"          ->  set_volume {"value": 40}
ENGLISH   "set gpio 17 high"              ->  gpio_on {"pin": 17}
ENGLISH   "shut down in 10 minutes"       ->  shutdown {"amount": 10, "unit": "min"}
HINGLISH  "volume 40 kar do"              ->  set_volume {"value": 40}
HINGLISH  "gpio 17 ko high karo"          ->  gpio_on {"pin": 17}
HINGLISH  "10 min baad shutdown kar dena" ->  shutdown {"amount": 10, "unit": "min"}

Model file	96.2 MB, FP32 PyTorch, 23,889,849 parameters
Runs on	CPU only, fully offline after download (Windows, Linux, Raspberry Pi 5)
Held-out test accuracy	94.87% exact (action + every argument) on 24,498 unseen-template commands; 96.05% action-only
Accuracy by language	ENGLISH 92.56% · HINGLISH 95.43% exact (8,038 and 9,958 held-out commands)
Speed	5.9 ms median per command on a laptop i7-1360P (2 threads); 40 ms median per command on a Raspberry Pi 5 (2 threads)
Links	GitHub code · HTML test guide · Hugging Face model

Executor notice. The model only parses commands. The executor included in this repository is a sample for testing, not a fully developed product. It is there so you can see a parsed command turn into a real action on your own machine. By default it refuses every critical action (shutdown, delete, run_command, GPIO on/off, service changes, ...). Use it on a test machine and read the plan before you run anything.

What it is for

The model is the "understanding" step of an on-device agent: speech-to-text (for example Whisper) → Union Command v10 → validator/risk gate → executor (tool call). It replaces a large LLM for the common, well-defined device commands, so the reply is instant, private and free.

Every use case is shown in ENGLISH first, then the same command in HINGLISH. The outputs are the real predictions of v10.

Use case	Language	Example command	Model output
Offline voice assistant for a laptop or home	ENGLISH	set the volume to 40	set_volume {"value": 40}
Raspberry Pi GPIO / I2C by voice or chat	ENGLISH	read the value of physical pin 11	gpio_read {"numbering": "board", "pin": 11}
Device diagnostics chat-ops	ENGLISH	what is the pi temperature	get_temperature {}
Cheap first-stage router for an LLM agent	ENGLISH	list docker containers	docker_list {}
Hands-free desktop and accessibility	ENGLISH	open notepad	open_app {"app": "notepad"}
Edge DevOps	ENGLISH	restart the nginx service	service_restart {"service": "nginx"}
Credentials copied verbatim	ENGLISH	connect to wifi Redmi Note 12 password hello@123	connect_wifi {"password": "hello@123", "ssid": "Redmi Note 12"}
Negated requests	ENGLISH	don't shut down the laptop	cancel_shutdown {}
Offline voice assistant for a laptop or home	HINGLISH	volume 40 kar do	set_volume {"value": 40}
Raspberry Pi GPIO / I2C by voice or chat	HINGLISH	physical pin 11 ki value padho	gpio_read {"numbering": "board", "pin": 11}
Device diagnostics chat-ops	HINGLISH	pi ka temperature batao	get_temperature {}
Cheap first-stage router for an LLM agent	HINGLISH	docker containers list karo	docker_list {}
Hands-free desktop and accessibility	HINGLISH	notepad kholo	open_app {"app": "notepad"}
Edge DevOps	HINGLISH	nginx service restart karo	service_restart {"service": "nginx"}
Credentials copied verbatim	HINGLISH	Redmi Note 12 wifi se connect karo password hello@123	connect_wifi {"password": "hello@123", "ssid": "Redmi Note 12"}
Negated requests	HINGLISH	shutdown mat karo	cancel_shutdown {}

For an LLM agent, send unknown or low-confidence text on to the bigger model. Vague commands such as turn it off come back as clarify, so the agent can ask the user what to switch off.

Every prediction returns JSON with the action, typed arguments, a confidence score, the top-3 alternative actions and schema validation errors:

{"action": "set_timer", "args": {"amount": 5, "unit": "min"}, "confidence": 0.9998,
  "alternatives": [{"action": "set_timer", "score": 0.9998}, ...], "validation_errors": [], "latency_ms": 6.1}

A good agent pattern: execute only when validation_errors is empty and confidence is high, ask the user when the action is clarify, and hand unknown or low-confidence text to a bigger model.

Categories, sample commands, outputs and executor tools

One intent per category: all ENGLISH commands first, then the same commands in HINGLISH. "Model output" is the real prediction of v10. "Executor tool" is what the sample executor would run for that output on each platform; the tag in brackets shows whether that adapter was ready on the test machine. "Result" is shown only for read-only commands that were actually executed; other rows were planned, not run.

Category	Language	Sample command	Model output	Executor tool: Windows 11	Executor tool: Raspberry Pi 5	Default gate	Result
Audio & media	ENGLISH	set the volume to 40	set_volume {"value": 40}	Python adapter executors/windows.py: Set master volume to an exact percent [needs pycaw]	Python adapter executors/common.py: Set master volume to an exact percent [needs pactl]	allowed	plan only (not executed in this demo)
Display & appearance	ENGLISH	lower the brightness a little	brightness_down {}	PowerShell: $b=Get-CimInstance -Namespace root/WMI -ClassName WmiMonitorBrightness… [ready]	Python adapter executors/common.py: Decrease brightness (default step 10) [needs brightnessctl]	allowed	plan only (not executed in this demo)
Network & connectivity	ENGLISH	how strong is the wifi signal	get_wifi_signal {}	netsh.exe wlan show interfaces [ready]	nmcli device wifi list [ready]	allowed	Windows: completed (output not published: local device details); Pi: completed (output not published: local device details)
GPIO & I2C	ENGLISH	set gpio 17 high	gpio_on {"pin": 17}	Python adapter executors/common.py: Set a GPIO pin HIGH (e.g. relay / LED on) [Pi only]	pinctrl set 17 op dh [ready]	blocked	blocked by default gate
Hardware & Raspberry Pi	ENGLISH	what is the pi temperature	get_temperature {}	PowerShell: $t=Get-CimInstance -Namespace root/wmi -ClassName MSAcpi_ThermalZoneTe… [ready]	vcgencmd measure_temp [ready]	allowed	Windows: failed: needs administrator rights; Pi: temp=43.9'C
Files & storage	ENGLISH	create file "notes/todo.txt"	create_file {"path": "notes/todo.txt"}	Python adapter tinyagent/executor.py: Create a new empty file; existing files are never overwritt… [ready]	Python adapter tinyagent/executor.py: Create a new empty file; existing files are never overwritt… [ready]	allowed	plan only (not executed in this demo)
Apps & windows	ENGLISH	open notepad	open_app {"app": "notepad"}	Python adapter executors/desktop.py: Open / launch an application [needs pywinauto]	Python adapter executors/desktop.py: Open / launch an application [ready]	allowed	plan only (not executed in this demo)
Browser & web	ENGLISH	search youtube for lofi music	youtube_search {"query": "lofi music"}	Python adapter executors/desktop.py: Search / play something on YouTube [ready]	Python adapter executors/desktop.py: Search / play something on YouTube [ready]	allowed	plan only (not executed in this demo)
Keyboard & clipboard	ENGLISH	select all	select_all {}	Python adapter executors/desktop.py: Send ctrl+a to the selected target window [needs pywinauto]	Python adapter executors/desktop.py: Send ctrl+a to the selected target window [needs wmctrl]	allowed	plan only (not executed in this demo)
Timers & productivity	ENGLISH	set a timer for 5 minutes	set_timer {"amount": 5, "unit": "min"}	Python adapter tinyagent/executor.py: Start a timer in this local server process [ready]	Python adapter tinyagent/executor.py: Start a timer in this local server process [ready]	allowed	plan only (not executed in this demo)
Security & accounts	ENGLISH	show the firewall status	firewall_status {}	PowerShell: Get-NetFirewallProfile | Select-Object Name,Enabled | ConvertTo-Json [ready]	Python adapter executors/common.py: show firewall status [needs ufw]	allowed	Windows: completed (output not published: local device details)
Services & processes	ENGLISH	show the status of the ssh service	service_status {"service": "ssh"}	PowerShell: Get-Service -Name $a.service | Select-Object Name,Status,DisplayName |… [ready]	systemctl status --no-pager -- ssh [ready]	allowed	Windows: failed: no such service on this OS
Software & development	ENGLISH	list docker containers	docker_list {}	docker.EXE ps -a [ready]	docker ps -a [ready]	allowed	Windows: completed (output not published: local device details); Pi: completed (output not published: local device details)
AI & models	ENGLISH	which models are available in ollama	list_llm_models {}	ollama.EXE list [ready]	ollama list [ready]	allowed	Windows: completed (output not published: local device details); Pi: completed (output not published: local device details)
Terminal & sessions	ENGLISH	show tmux sessions	list_sessions {}	Python adapter executors/common.py: List sessions [Pi only]	Python adapter executors/common.py: List sessions [needs tmux]	allowed	plan only (not executed in this demo)
System & power	ENGLISH	what is the cpu usage	get_cpu_usage {}	Python adapter tinyagent/executor.py: Measure CPU use over 200 milliseconds [ready]	Python adapter tinyagent/executor.py: Measure CPU use over 200 milliseconds [ready]	allowed	Windows: CPU 7.5 %; Pi: CPU 0.0 %
Help & intent handling	ENGLISH	what can you do	help {}	Python adapter tinyagent/executor.py: Show action and adapter coverage [ready]	Python adapter tinyagent/executor.py: Show action and adapter coverage [ready]	allowed	Windows: 371 actions, 221 live adapters here; Pi: 371 actions, 251 live adapters here
Audio & media	HINGLISH	volume 40 kar do	set_volume {"value": 40}	Python adapter executors/windows.py: Set master volume to an exact percent [needs pycaw]	Python adapter executors/common.py: Set master volume to an exact percent [needs pactl]	allowed	plan only (not executed in this demo)
Display & appearance	HINGLISH	brightness thoda kam karo	brightness_down {}	PowerShell: $b=Get-CimInstance -Namespace root/WMI -ClassName WmiMonitorBrightness… [ready]	Python adapter executors/common.py: Decrease brightness (default step 10) [needs brightnessctl]	allowed	plan only (not executed in this demo)
Network & connectivity	HINGLISH	wifi ka signal kitna strong hai	get_wifi_signal {}	netsh.exe wlan show interfaces [ready]	nmcli device wifi list [ready]	allowed	Windows: completed (output not published: local device details); Pi: completed (output not published: local device details)
GPIO & I2C	HINGLISH	gpio 17 ko high karo	gpio_on {"pin": 17}	Python adapter executors/common.py: Set a GPIO pin HIGH (e.g. relay / LED on) [Pi only]	pinctrl set 17 op dh [ready]	blocked	blocked by default gate
Hardware & Raspberry Pi	HINGLISH	pi ka temperature batao	get_temperature {}	PowerShell: $t=Get-CimInstance -Namespace root/wmi -ClassName MSAcpi_ThermalZoneTe… [ready]	vcgencmd measure_temp [ready]	allowed	Windows: failed: needs administrator rights; Pi: temp=43.3'C
Files & storage	HINGLISH	file "notes/todo.txt" bana do	create_file {"path": "notes/todo.txt"}	Python adapter tinyagent/executor.py: Create a new empty file; existing files are never overwritt… [ready]	Python adapter tinyagent/executor.py: Create a new empty file; existing files are never overwritt… [ready]	allowed	plan only (not executed in this demo)
Apps & windows	HINGLISH	notepad kholo	open_app {"app": "notepad"}	Python adapter executors/desktop.py: Open / launch an application [needs pywinauto]	Python adapter executors/desktop.py: Open / launch an application [ready]	allowed	plan only (not executed in this demo)
Browser & web	HINGLISH	youtube pe lofi music search karo	youtube_search {"query": "lofi music"}	Python adapter executors/desktop.py: Search / play something on YouTube [ready]	Python adapter executors/desktop.py: Search / play something on YouTube [ready]	allowed	plan only (not executed in this demo)
Keyboard & clipboard	HINGLISH	sab select karo	select_all {}	Python adapter executors/desktop.py: Send ctrl+a to the selected target window [needs pywinauto]	Python adapter executors/desktop.py: Send ctrl+a to the selected target window [needs wmctrl]	allowed	plan only (not executed in this demo)
Timers & productivity	HINGLISH	5 minute ka timer lagao	set_timer {"amount": 5, "unit": "min"}	Python adapter tinyagent/executor.py: Start a timer in this local server process [ready]	Python adapter tinyagent/executor.py: Start a timer in this local server process [ready]	allowed	plan only (not executed in this demo)
Security & accounts	HINGLISH	firewall ka status batao	firewall_status {}	PowerShell: Get-NetFirewallProfile | Select-Object Name,Enabled | ConvertTo-Json [ready]	Python adapter executors/common.py: show firewall status [needs ufw]	allowed	Windows: completed (output not published: local device details)
Services & processes	HINGLISH	ssh service ka status dikhao	service_status {"service": "ssh"}	PowerShell: Get-Service -Name $a.service | Select-Object Name,Status,DisplayName |… [ready]	systemctl status --no-pager -- ssh [ready]	allowed	Windows: failed: no such service on this OS
Software & development	HINGLISH	docker containers list karo	docker_list {}	docker.EXE ps -a [ready]	docker ps -a [ready]	allowed	Windows: completed (output not published: local device details); Pi: completed (output not published: local device details)
AI & models	HINGLISH	ollama pe kaunse models hain	list_llm_models {}	ollama.EXE list [ready]	ollama list [ready]	allowed	Windows: completed (output not published: local device details); Pi: completed (output not published: local device details)
Terminal & sessions	HINGLISH	tmux sessions dikhao	list_sessions {}	Python adapter executors/common.py: List sessions [Pi only]	Python adapter executors/common.py: List sessions [needs tmux]	allowed	plan only (not executed in this demo)
System & power	HINGLISH	cpu usage batao	get_cpu_usage {}	Python adapter tinyagent/executor.py: Measure CPU use over 200 milliseconds [ready]	Python adapter tinyagent/executor.py: Measure CPU use over 200 milliseconds [ready]	allowed	Windows: CPU 6.7 %; Pi: CPU 3.7 %
Help & intent handling	HINGLISH	tum kya kya kar sakte ho	help {}	Python adapter tinyagent/executor.py: Show action and adapter coverage [ready]	Python adapter tinyagent/executor.py: Show action and adapter coverage [ready]	allowed	Windows: 371 actions, 221 live adapters here; Pi: 371 actions, 251 live adapters here

All 72 example commands: 36 intents, each in ENGLISH and HINGLISH

Category	Language	Command	Model output	Confidence	Correct
Audio & media	ENGLISH	set the volume to 40	set_volume {"value": 40}	100%	✓
Audio & media	ENGLISH	pause the music	media_pause {}	100%	✓
Display & appearance	ENGLISH	lower the brightness a little	brightness_down {}	100%	✓
Display & appearance	ENGLISH	turn on dark mode	dark_mode_on {}	100%	✓
Network & connectivity	ENGLISH	how strong is the wifi signal	get_wifi_signal {}	100%	✓
Network & connectivity	ENGLISH	connect to wifi Redmi Note 12 password hello@123	connect_wifi {"password": "hello@123", "ssid": "Redmi Note 12"}	100%	✓
GPIO & I2C	ENGLISH	set gpio 17 high	gpio_on {"pin": 17}	100%	✓
GPIO & I2C	ENGLISH	read the value of physical pin 11	gpio_read {"numbering": "board", "pin": 11}	100%	✓
GPIO & I2C	ENGLISH	scan the i2c bus	i2c_scan {}	100%	✓
Hardware & Raspberry Pi	ENGLISH	what is the pi temperature	get_temperature {}	100%	✓
Hardware & Raspberry Pi	ENGLISH	what is the fan speed	get_fan_speed {}	100%	✓
Files & storage	ENGLISH	create file "notes/todo.txt"	create_file {"path": "notes/todo.txt"}	100%	✓
Files & storage	ENGLISH	how much disk space is left	get_disk_space {}	100%	✓
Apps & windows	ENGLISH	open notepad	open_app {"app": "notepad"}	100%	✓
Apps & windows	ENGLISH	minimize this window	minimize_window {}	100%	✓
Browser & web	ENGLISH	search youtube for lofi music	youtube_search {"query": "lofi music"}	100%	✓
Browser & web	ENGLISH	login to github username dev_user password Test@123	web_login {"password": "Test@123", "site": "github", "username": "dev_user"}	100%	✓
Keyboard & clipboard	ENGLISH	select all	select_all {}	100%	✓
Keyboard & clipboard	ENGLISH	show clipboard history	clipboard_history {}	100%	✓
Timers & productivity	ENGLISH	set a timer for 5 minutes	set_timer {"amount": 5, "unit": "min"}	100%	✓
Timers & productivity	ENGLISH	calculate 12 + 8	calculate {"expression": "12 + 8"}	100%	✓
Security & accounts	ENGLISH	show the firewall status	firewall_status {}	100%	✓
Security & accounts	ENGLISH	allow port 8080 in the firewall	firewall_allow {"port": 8080}	100%	✓
Services & processes	ENGLISH	show the status of the ssh service	service_status {"service": "ssh"}	100%	✓
Services & processes	ENGLISH	restart the nginx service	service_restart {"service": "nginx"}	100%	✓
Software & development	ENGLISH	list docker containers	docker_list {}	100%	✓
Software & development	ENGLISH	install the numpy python package	install_python_package {"name": "numpy"}	100%	✓
AI & models	ENGLISH	which models are available in ollama	list_llm_models {}	100%	✓
AI & models	ENGLISH	run the llama3 model	run_llm {"model": "llama3"}	100%	✓
Terminal & sessions	ENGLISH	show tmux sessions	list_sessions {}	100%	✓
Terminal & sessions	ENGLISH	run command "ls -la"	run_command {"command": "command \"ls -la"}	100%	✗ (see note)
System & power	ENGLISH	what is the cpu usage	get_cpu_usage {}	100%	✓
System & power	ENGLISH	shut down in 10 minutes	shutdown {"amount": 10, "unit": "min"}	100%	✓
System & power	ENGLISH	don't shut down the laptop	cancel_shutdown {}	100%	✓
Help & intent handling	ENGLISH	what can you do	help {}	99%	✓
Help & intent handling	ENGLISH	turn it off	clarify {}	100%	✓
Audio & media	HINGLISH	volume 40 kar do	set_volume {"value": 40}	100%	✓
Audio & media	HINGLISH	gaana pause karo	media_pause {}	100%	✓
Display & appearance	HINGLISH	brightness thoda kam karo	brightness_down {}	100%	✓
Display & appearance	HINGLISH	dark mode on kar do	dark_mode_on {}	100%	✓
Network & connectivity	HINGLISH	wifi ka signal kitna strong hai	get_wifi_signal {}	100%	✓
Network & connectivity	HINGLISH	Redmi Note 12 wifi se connect karo password hello@123	connect_wifi {"password": "hello@123", "ssid": "Redmi Note 12"}	100%	✓
GPIO & I2C	HINGLISH	gpio 17 ko high karo	gpio_on {"pin": 17}	100%	✓
GPIO & I2C	HINGLISH	physical pin 11 ki value padho	gpio_read {"numbering": "board", "pin": 11}	100%	✓
GPIO & I2C	HINGLISH	i2c bus scan karo	i2c_scan {}	100%	✓
Hardware & Raspberry Pi	HINGLISH	pi ka temperature batao	get_temperature {}	82%	✓
Hardware & Raspberry Pi	HINGLISH	fan speed kitni hai	get_fan_speed {}	100%	✓
Files & storage	HINGLISH	file "notes/todo.txt" bana do	create_file {"path": "notes/todo.txt"}	100%	✓
Files & storage	HINGLISH	disk space kitna bacha hai	get_disk_space {}	100%	✓
Apps & windows	HINGLISH	notepad kholo	open_app {"app": "notepad"}	100%	✓
Apps & windows	HINGLISH	is window ko minimize karo	minimize_window {}	100%	✓
Browser & web	HINGLISH	youtube pe lofi music search karo	youtube_search {"query": "lofi music"}	100%	✓
Browser & web	HINGLISH	github pe login karo username dev_user password Test@123	web_login {"password": "Test@123", "site": "github", "username": "dev_user"}	100%	✓
Keyboard & clipboard	HINGLISH	sab select karo	select_all {}	100%	✓
Keyboard & clipboard	HINGLISH	clipboard history dikhao	clipboard_history {}	100%	✓
Timers & productivity	HINGLISH	5 minute ka timer lagao	set_timer {"amount": 5, "unit": "min"}	100%	✓
Timers & productivity	HINGLISH	12 + 8 calculate karo	calculate {"expression": "12 + 8"}	100%	✓
Security & accounts	HINGLISH	firewall ka status batao	firewall_status {}	100%	✓
Security & accounts	HINGLISH	firewall mein port 8080 allow karo	firewall_allow {"port": 8080}	100%	✓
Services & processes	HINGLISH	ssh service ka status dikhao	service_status {"service": "ssh"}	100%	✓
Services & processes	HINGLISH	nginx service restart karo	service_restart {"service": "nginx"}	100%	✓
Software & development	HINGLISH	docker containers list karo	docker_list {}	100%	✓
Software & development	HINGLISH	numpy python package install karo	install_python_package {"name": "numpy"}	100%	✓
AI & models	HINGLISH	ollama pe kaunse models hain	list_llm_models {}	100%	✓
AI & models	HINGLISH	llama3 model chalao	run_llm {"model": "llama3"}	100%	✓
Terminal & sessions	HINGLISH	tmux sessions dikhao	list_sessions {}	100%	✓
Terminal & sessions	HINGLISH	command "ls -la" chalao	run_command {"command": "command \"ls -la"}	100%	✗ (see note)
System & power	HINGLISH	cpu usage batao	get_cpu_usage {}	100%	✓
System & power	HINGLISH	10 min baad shutdown kar dena	shutdown {"amount": 10, "unit": "min"}	100%	✓
System & power	HINGLISH	shutdown mat karo	cancel_shutdown {}	100%	✓
Help & intent handling	HINGLISH	tum kya kya kar sakte ho	help {}	100%	✓
Help & intent handling	HINGLISH	ise band kar do	close_window {}	100%	✗ (see note)

All 72 illustrative commands are in examples/sample_commands.jsonl, with per-platform executor plans in examples/category_examples_windows.json and examples/category_examples_pi.json. v10 got 69 of 72 exactly right (ENGLISH 35 of 36, HINGLISH 34 of 36). The misses:

• ENGLISH run command "ls -la" gave run_command {"command": "command \"ls -la"}; expected run_command {"command": "ls -la"}.
• HINGLISH command "ls -la" chalao gave run_command {"command": "command \"ls -la"}; expected run_command {"command": "ls -la"}.
• HINGLISH ise band kar do gave close_window {}; expected clarify {}.

These commands were written for this demo, so they are illustrative. The accuracy numbers below come from the held-out test set.

Test results

Accuracy

Metric: exact match, meaning the action and every typed argument must be identical to the label. Errors and abstentions stay in the denominator. Same checkpoint on both devices: SHA-256 c91d44e4a687152cd65d86d56cfc0e453a60a1afe14f8c395e8af058f2fc4d5b.

Dataset	Rows	Laptop CPU exact	Laptop action-only	Raspberry Pi 5 exact	Pi action-only	About the set
Held-out test (unseen templates)	24,498	94.87% (23,241)	96.05%	94.87% (23,241)	96.05%	24,498 rows · all 371 actions · template families never seen in training
Practical dev (A)	1,321	96.44% (1,274)	96.90%	96.44% (1,274)	96.90%	1,321 realistic phrasings · used during development
Golden dev (B)	383	97.39% (373)	97.91%	97.39% (373)	97.91%	383 hand-checked commands · used during development

The held-out test split is made of whole template families that were never used for training. It was evaluated once for this release. Practical dev and golden dev guided earlier development, so treat them as regression checks. Earlier development measurements on the same checkpoint (Windows CPU): Grouped dev (reserved template groups) 96.50% (30,596/31,707), Validation 98.22% (23,467/23,892).

Held-out test by language:

Language	Rows	Exact	Action-only
HINGLISH (Roman script, incl. a small Devanagari slice)	9,958	95.43%	96.37%
ENGLISH	8,038	92.56%	94.31%
Unlabelled (project A rows, mostly English/Hinglish)	4,884	98.73%	99.45%
Code-mixed	1,618	91.22%	92.46%

Held-out test by category:

Category	Rows	Exact	Action-only
Hardware & Raspberry Pi	1,122	98.93%	98.93%
Terminal & sessions	410	98.78%	98.78%
Display & appearance	1,972	98.48%	98.53%
Audio & media	1,653	98.12%	98.31%
Browser & web	1,527	97.84%	99.61%
System & power	3,085	97.50%	98.64%
Security & accounts	714	96.78%	96.78%
Keyboard & clipboard	883	96.60%	97.28%
Files & storage	2,793	95.31%	96.89%
Help & intent handling	762	95.14%	96.19%
Apps & windows	1,468	94.96%	96.66%
Services & processes	1,088	93.57%	94.76%
Software & development	926	92.33%	92.44%
Network & connectivity	2,951	91.70%	92.82%
Timers & productivity	747	90.63%	93.71%
GPIO & I2C	2,133	87.48%	88.70%
AI & models	264	80.68%	97.73%

Most frequent action confusions on the held-out test:

Expected	Predicted	Count
gpio_stop_watch	gpio_off	59
i2c_read	i2c_write	54
create_service	service_enable	52
hotspot_on	unknown	51
get_public_ip	get_ip	51
scan_lan	discover_devices	37
gpio_blink	gpio_pulse	35
enable_auto_updates	install_python_package	35

Speed

Warm, single-command latency over one command per catalog action (371 commands, 2 passes), then batched throughput (batch 32). Includes Python and PyTorch overhead. "RAM" is the whole process after loading the model.

Device	Threads	Median	p95	Single commands/s	Batched commands/s	Model load	RAM
Laptop i7-1360P	1	7.4 ms	9.9 ms	131	141	5.3 s	650 MB
Laptop i7-1360P	2	5.9 ms	9.4 ms	157	249	5.3 s	650 MB
Laptop i7-1360P	4	7.2 ms	11.3 ms	128	237	5.3 s	650 MB
Raspberry Pi 5	1	32.8 ms	40.8 ms	30	91	8.3 s	533 MB
Raspberry Pi 5	2	40.4 ms	47.6 ms	24	88	8.3 s	533 MB
Raspberry Pi 5	4	36.2 ms	44.0 ms	27	100	8.3 s	533 MB

Full accuracy run speed (batched, held-out test):

Device	Seconds	Commands/s	Threads	Batch
Laptop i7-1360P	95.86	255.5	4	64
Raspberry Pi 5	331.81	73.8	2	64

Temperature

Device	Run	Sensor	Start	Max	Mean	Pi throttle flags after run
Laptop i7-1360P	Accuracy run (26k rows, 4 threads)	Windows ACPI thermal zone (typeperf, Thermal Zone Information)	60.9 °C	95.9 °C	85.3 °C	n/a (not a Pi)
Laptop i7-1360P	Latency benchmark (1/2/4 threads)	Windows ACPI thermal zone (typeperf, Thermal Zone Information)	53.9 °C	85.9 °C	71.2 °C	n/a (not a Pi)
Raspberry Pi 5	Accuracy run	sysfs cpu-thermal (Raspberry Pi SoC sensor)	51.8 °C	73.2 °C	68.2 °C	throttled=0x0
Raspberry Pi 5	Latency benchmark	sysfs cpu-thermal (Raspberry Pi SoC sensor)	41.9 °C	65.5 °C	56.5 °C	throttled=0x0

Raspberry Pi 5 notes. The first attempt ran the held-out test in one go with 4 threads. After 202 s and 15,040 commands the SoC reached about 85 °C with the fan near 9,900 RPM, and the Pi dropped off the network until it was power-cycled. The reported Pi results come from a second, gentler run: 2 threads, the held-out test split into five chunks of 5,000 rows, and a cool-down below 65 °C before each chunk. Temperature, the 5 V rail, throttle flags and fan speed were logged every 5 seconds (test_results/rpi5/sensors.log). That run finished with throttle flags at 0x0 throughout, a peak of 73.2 °C and a 5 V rail between 5.05 V and 5.20 V. The Pi made exactly the same predictions as the laptop: the same 1,314 commands were wrong on both devices across the three sets. On the Pi, 1 thread gives the lowest single-command latency; extra threads mainly help batched throughput. For long jobs on a Pi 5, use 1 or 2 threads, keep the active cooler and a 5 V / 5 A supply, and watch vcgencmd measure_temp and vcgencmd get_throttled.

Test devices

Device	CPU	OS	Python / PyTorch
Laptop	13th Gen Intel(R) Core(TM) i7-1360P	Windows-11-10.0.26200-SP0	Python 3.13.14 / torch 2.12.0+cpu
Raspberry Pi 5	Raspberry Pi 5 Model B Rev 1.1 (Arm Cortex-A76, 4 cores)	Linux-6.12.47+rpt-rpi-2712-aarch64-with-glibc2.41	Python 3.13.5 / torch 2.8.0+cpu

Download and test

1. Get the code and the model

git clone https://github.com/sraivante/tiny-agentic-home-robotic-for-edge-device-v10.git
cd tiny-agentic-home-robotic-for-edge-device-v10
python download_model.py      # 96.2 MB from Hugging Face, SHA-256 verified

Or download the ZIP of the repository. The model file alone is at sraivante/tiny-agentic-home-robotic-for-edge-device-v10.

2. Install (Python 3.11 to 3.13, CPU only)

Windows (PowerShell):

python -m venv .venv
.venv\Scripts\activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt

Raspberry Pi OS 64-bit / Linux:

python3 -m venv .venv
. .venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt

3. Parse commands (nothing is executed)

python quickstart.py                                  # demo: ENGLISH first, then HINGLISH
python quickstart.py "turn off the wifi" "wifi band karo"
python quickstart.py --json "set a timer for 5 minutes"

From Python:

from tinyagent.runtime import Predictor
model = Predictor("models/a100_minilm_v10_quoted/best.pt", threads=2)
print(model.predict("set the volume to 40"))    # ENGLISH
print(model.predict("volume 40 kar do"))        # HINGLISH

4. Reproduce the tests on your device

python scripts/evaluate.py --data examples/sample_commands.jsonl --output my_eval.json
python scripts/benchmark.py --threads 1,2,4 --output my_benchmark.json

evaluate.py accepts any JSONL file with text, action and args. The full training and test datasets are not published; the measured reports are in test_results/.

5. Try the sample executor lab (optional)

run.bat              # Windows
bash run.sh          # Linux / Raspberry Pi

The launcher installs what it needs into .venv, downloads the model if missing and opens http://127.0.0.1:8770. Pick a category and an example (or type your own), then press Predict only to see the parse, or Run & execute to let the sample executor act on this computer.

• It listens on 127.0.0.1 only and needs a per-session token for every request.
• File actions are confined to runtime/files/.
• critical actions are refused unless you start it with --allow-risk critical.
• Many adapters need extra tools or hardware (pycaw, pywinauto, pinctrl, i2c-tools, NetworkManager, Docker, Ollama). The UI shows which ones are ready. See executor_config.example.json for the integrations that need your own settings.

The executor is a sample for testing, not a finished product. Adapter coverage, error handling and security hardening are incomplete. Do not expose it to a network or run it unattended.

How the model works

• Inputs: one command, up to 256 characters.
• Encoders: a character-level Transformer (width 128, 3 layers) for typos and code-mixing, plus the MiniLM-L6 word encoder from sentence-transformers/all-MiniLM-L6-v2 (fine-tuned), fused per character.
• Heads: a 371-way action classifier and typed argument heads. Each argument is either copied as an exact span of the input (names, paths, passwords, numbers) or chosen from a learned label set (units, modes, apps).
• Decoding v3: keeps quoted file names, host names and whole numbers intact when a span overlaps them.
• Validation: every output is checked against the catalog (union_catalog.json): required slots, types, ranges such as volume 0 to 100 and BCM pin 0 to 27.
• Training: 1,143,182 synthetic fitting rows, NVIDIA A100 (bf16), epoch 5 of 9 selected by grouped-dev accuracy.
• Catalog metadata: each action carries risk (safe/caution/critical), needs_sudo and platforms, so your executor can decide what to run, confirm or refuse. The model never refuses anything itself.

Limitations

• Training data is synthetic. There is no real speech-to-text output in it, so test with your own ASR transcripts.
• Devanagari Hindi coverage is small. English and Roman-script Hinglish dominate.
• One command per utterance. It is a parser, not a multi-step planner.
• Weakest areas on the held-out test: near-duplicate GPIO/I2C actions (i2c_read vs i2c_write, gpio_blink vs gpio_pulse), a few near-synonyms (get_public_ip vs get_ip, scan_lan vs discover_devices) and some free-text argument boundaries, like the run_command example above.
• The model does not refuse dangerous requests. Always put a validator and a risk gate in front of any executor.

Files

Path	What it is
models/a100_minilm_v10_quoted/best.pt	The model (Hugging Face only; GitHub users run download_model.py)
config.json	Model metadata; also the file Hugging Face uses to count downloads
tinyagent/	Model and inference code (Predictor) plus the base executor
union_catalog.json	The 371 actions with slots, risk, sudo and platform metadata
quickstart.py, download_model.py	Command-line demo and verified downloader
scripts/	evaluate.py, benchmark.py, category_examples.py, build_docs.py
app.py, lab_executor.py, executors/, web/	The sample executor lab
test_results/	Accuracy, speed and temperature reports for the laptop and the Raspberry Pi 5
examples/	Illustrative commands and per-platform executor plans
docs/index.html	The HTML test guide (online)

License and credits

Code and model weights: Apache-2.0, Copyright (c) 2026 sraivante. The word encoder is initialised from sentence-transformers/all-MiniLM-L6-v2 (Apache-2.0); see provenance/base_models/minilm_l6/. The training dataset is synthetic and is not published. Provided as-is, without warranty. You are responsible for anything you let an executor do.