Da-The-Pashto-Empowerment-GGUF

A Pashto reasoning model built from nassimjp/MiniCPM5-1B-Pashto, converted to GGUF for local inference.

Da-The-Pashto-Empowerment is part of the iPashto.ai effort to improve Pashto-language AI through open models, specialized tokenization, reasoning data, and efficient local deployment.

This repository contains GGUF versions of the fine-tuned model for use with llama.cpp and compatible local inference applications.

Model Overview

Model lineage

openbmb/MiniCPM5-1B
          |
          v
nassimjp/MiniCPM5-1B-Pashto
          |
          | Pashto/Urdu tokenizer modification
          | 46 additional single-token characters
          v
Da-The-Pashto-Empowerment
          |
          | Reasoning / Thinking fine-tuning
          v
Da-The-Pashto-Empowerment-GGUF
          |
          +-- F16
          +-- Q8_0
          +-- Q6_K
          +-- Q4_K_M
          +-- Q3_K_M

The parent model introduces 46 additional Pashto- and Urdu-relevant characters as individual tokenizer tokens. This child model builds on that parent through reasoning-oriented fine-tuning and GGUF conversion.

The diagram describes the intended model lineage. Refer to the parent model card for the tokenizer modification details.

Key Features

• Pashto-focused AI: Developed to support Pashto-language text generation and reasoning.
• Reasoning-oriented fine-tuning: Trained using reasoning and thinking examples.
• Tokenizer-aware foundation: Inherits its model lineage from the Pashto/Urdu tokenizer-modified parent.
• Multiple GGUF quantizations: Choose between higher numerical precision and smaller storage requirements.
• Local inference: Designed for compatible local runtimes, including llama.cpp.
• Open model access: Published on Hugging Face for experimentation, evaluation, and further development.

The model is intended for experimentation and research. Its reasoning ability, Pashto fluency, factual accuracy, and multilingual behavior should be evaluated independently.

Training Summary

The following figures describe the reported training run.

Important: Training loss is an optimization metric, not a direct measure of reasoning accuracy. These figures do not independently establish benchmark performance or superiority over the parent model.

The model was exported to GGUF using Unsloth.

Available GGUF Files

Five model variants are provided.

File sizes are approximate and may vary slightly with repository revisions.

Which version should you choose?

• F16: Use when you want a higher-precision reference and have sufficient memory.
• Q8_0: A good option when you want high precision with a smaller footprint than F16.
• Q6_K: A useful middle ground between size and precision.
• Q4_K_M: A practical starting point for many local deployments.
• Q3_K_M: Choose when storage or memory is especially limited; lower-bit quantization can reduce output quality.

Quantization does not guarantee identical results across variants. Test the versions on your own Pashto evaluation set.

Quick Start with llama.cpp

1. Install llama.cpp

Follow the official installation instructions:

https://github.com/ggml-org/llama.cpp

Make sure your installation supports the model architecture and GGUF format used by this repository.

2. Run directly from Hugging Face

For a text-only model, try:

llama-cli \
  -hf nassimjp/Da-The-Pashto-Empowerment-GGUF \
  --jinja

To request a specific quantization, use the repository's supported Hugging Face quantization selector. For example, if the selector resolves the corresponding file in this repository:

llama-cli \
  -hf nassimjp/Da-The-Pashto-Empowerment-GGUF:Q4_K_M \
  --jinja

If automatic file selection is unavailable in your installed version, download the desired GGUF file and load it directly.

3. Run a downloaded GGUF file

Example using Q4_K_M:

llama-cli \
  -m MiniCPM5-1B-Pashto.Q4_K_M.gguf \
  --jinja \
  -n 512 \
  --temp 0.7

Replace the filename with the exact file you downloaded.

The values above are example generation settings, not a guaranteed optimal configuration for every task.

4. Start a local server

llama-server \
  -m MiniCPM5-1B-Pashto.Q4_K_M.gguf \
  --jinja \
  --host 127.0.0.1 \
  --port 8080 \
  -c 8192

The server provides a local HTTP interface when supported by your llama.cpp build.

You can then connect a compatible local chat interface or application to:

http://127.0.0.1:8080

The context length in this example is a configurable setting, not a claim that every deployment can run that context size within its available memory.

Example Pashto Prompt

پوښتنه:
یو بزګر ۱۷ پسونه لري. ټول مري، پرته له ۹ څخه.
څو پسونه ورته پاتې شول؟

لارښوونه:
پوښتنه په دقت ولوله، د جملې مانا وڅېړه،
او ځواب په طبیعي پښتو کې له لنډ دلیل سره ولیکه.

Expected answer:

۹ پسونه ورته پاتې شول، ځکه په جمله کې ویل شوي
چې ټول پسونه مري، پرته له ۹ څخه.

This is an example evaluation prompt, not a reported benchmark result.

Evaluation Recommendations

For meaningful evaluation, test the fine-tuned model against its parent model using the same prompts and comparable generation settings.

Recommended evaluation categories:

1. Pashto fluency: Grammar, spelling, natural phrasing, and vocabulary.
2. Logical reasoning: Multi-step problems, word problems, and reasoning puzzles.
3. Instruction following: Whether the model answers the actual question.
4. Factual accuracy: Whether it distinguishes established facts from assumptions.
5. Consistency: Whether the conclusion follows from the explanation.
6. Multilingual contamination: Whether unrelated languages appear in Pashto responses.
7. Quantization impact: Whether lower-bit variants retain useful reasoning performance.

Keep evaluation questions separate from training data whenever possible. Report accuracy, failure cases, generation settings, and hardware alongside any benchmark claims.

No standardized benchmark scores are claimed in this model card.

Limitations

• A model of approximately 1B parameters can struggle with complex reasoning and factual recall.
• Reasoning-oriented fine-tuning does not guarantee correct reasoning on every question.
• The model may produce unsupported claims, contradictions, repetition, or grammatical errors.
• Performance can vary with the prompt template, context length, sampling parameters, and inference runtime.
• Quantization can affect output quality.
• Multilingual support in the model lineage does not guarantee equal capability in Pashto, Urdu, Persian, Arabic, and English.
• Tokenizer modifications improve character representation but do not, by themselves, teach the model the language.

Please verify important outputs before using them in educational, professional, or other high-stakes settings.

License and Attribution

This repository declares the Apache-2.0 license. Users should also review the upstream model's license and comply with all applicable license and attribution requirements.

• Fine-tuned parent: nassimjp/MiniCPM5-1B-Pashto
• Original upstream model: openbmb/MiniCPM5-1B
• GGUF conversion tool: Unsloth
• Inference runtime: llama.cpp

Acknowledgements

Thanks to the MiniCPM and OpenBMB developers, the Unsloth team, the llama.cpp community, Hugging Face, and the open-source language technology community.

This work is part of the broader effort to make Pashto-language AI models more accessible, practical, and useful for researchers and everyday users.

Created for Pashto AI empowerment.

No Unicode character left behind!