README.md

metadata

title: ZamAI mT5 Pashto Demo
emoji: 🤖
colorFrom: green
colorTo: red
sdk: gradio
sdk_version: 5.43.1
app_file: app.py
pinned: false

ZamAI mT5 Pashto Instruction-Following Demo

This Space showcases the tasal9/ZamAI-mT5-Pashto model, a fine-tuned version of google/mt5-base for instruction-following in the Pashto language.

How to Use

1. Enter an Instruction: Provide a command or task in the "Instruction" textbox.
2. (Optional) Provide Input: If the instruction requires additional context, add it to the "Input" textbox.
3. Adjust Settings: You can modify generation parameters like max_length, num_beams, temperature, and top_p using the sliders.
4. Generate: Click the "Generate" button to see the model's response.

About the Model

This model was fine-tuned on the tasal9/ZamAI-Pashto-High-Qualituly-Dataset. It is designed to understand and execute a wide variety of instructions in Pashto.

Modes (Runtime Behavior)

The UI exposes a Mode dropdown:

• off – Real model generation (loads weights, may take time first run)
• echo – Returns the prompt exactly (fast, for UI testing)
• useless – Returns a fixed placeholder response (for latency / wiring checks)

You can also preset a default via environment: ECHO_MODE=echo or ECHO_MODE=useless before launching. The UI selection overrides the env per request.

Running Locally

python app.py

Environment variables (optional):

Variable	Purpose	Example
MODEL_ID	Override model repo	tasal9/ZamAI-mT5-Pashto
GRADIO_PORT	Change port	9000
GRADIO_HOST	Bind address	0.0.0.0
ECHO_MODE	Default mode (off|echo|useless)	echo
HF_HOME	Hugging Face cache root	/data/hf-cache
OFFLINE	If set to 1/true, sets HF_HUB_OFFLINE=1	1
DEFAULT_MAX_NEW_TOKENS	Slider default tokens	128

Example:

HF_HOME=/data/hf-cache OFFLINE=1 ECHO_MODE=echo python app.py

Caching & Offline Usage

1. Pick a persistent cache directory (fast disk/SSD recommended).
2. Set HF_HOME to that path on every run.
3. Launch once in real mode (off) so weights download.
4. Set OFFLINE=1 (or directly HF_HUB_OFFLINE=1) for subsequent offline usage.

Linux example:

export HF_HOME=/data/hf-cache
python app.py  # first run downloads
export OFFLINE=1
python app.py  # offline reuse

Windows PowerShell example:

$env:HF_HOME = 'D:\hf-cache'
python app.py   # first run
$env:OFFLINE = '1'
python app.py   # offline

Health Endpoint

A lightweight HTTP health check server listens on HEALTH_PORT (default 8080).

curl http://localhost:8080/health  # returns 'ok'

Development Tips

• Use echo mode to iterate on UI changes without loading the full model.
• useless mode provides a constant response for benchmarking layout latency.
• Logs show cache configuration: HF_HOME, TRANSFORMERS_CACHE, and offline status.
• GPU is auto-detected (CUDA) and displayed in the header if available.

Common Issues

Symptom	Likely Cause	Fix
Slow first generation	Model weights not cached	Ensure persistent HF_HOME
Network calls while offline	Missing OFFLINE=1	Export OFFLINE=1 (sets HF_HUB_OFFLINE)
IndentationError on Space	Space not rebuilt	Trigger rebuild / push a new commit
Echo mode still loads model	Open runs before switching	Set ECHO_MODE=echo before launch

Trigger Rebuild (Hugging Face Space)

If a rebuild doesn’t trigger automatically, make a no-op commit:

date > .rebuild
git add .rebuild
git commit -m "chore: trigger rebuild"
git push origin main

License / Usage

Refer to the model card of tasal9/ZamAI-mT5-Pashto for any licensing or usage constraints. This demo adds only interface logic and does not modify model licensing terms.