nassimjp/Pashto-Brain-Extraction-Dataset

🧠 Pashto Brain Extraction Dataset

A small experimental Pashto reasoning dataset designed to extract and preserve useful model reasoning/planning while discarding the final answer.

Keep the brain 🧠 — throw away the mouth 🗣️

🎯 Purpose

A language model may understand a question and produce useful reasoning while still generating poor, unnatural, or grammatically incorrect Pashto in its final answer.

Instead of throwing away the entire generation, this dataset preserves the reasoning and removes the final answer.

🧠 Core Concept

Question
   ↓
Teacher Model
   ↓
Reasoning / Brain 🧠
   ↓
Final Answer ❌ discarded
   ↓
Pashto Writer Model
   ↓
Natural Pashto Answer ✅

The main idea is:

Brain ≠ Mouth

Different models can have different strengths. One model may provide useful reasoning while another model may be much better at producing natural Pashto.

📦 Dataset Format

The dataset uses the conversational messages format.

Each record contains a user question and an assistant reasoning block:

{
  "messages": [
    {
      "role": "user",
      "content": "یوه نجلۍ د بازار په ګڼه ګوڼه کې له خپلې مور څخه جلا شوه؛ څنګه یې خپله مور پیدا کړه؟"
    },
    {
      "role": "assistant",
      "content": "<think>\nنجلۍ د بازار په ګڼه ګوڼه کې له مور څخه جلا کېږي. هغه باید په خپل ځای کې ودرېږي، د یوه باوري کس څخه مرسته وغواړي او د بازار له ساتونکي سره اړیکه ونیسي، تر څو خپله مور پیدا کړي.\n</think>\n\n"
    }
  ]
}

Important

The final answer is intentionally empty.

Only the reasoning is preserved:

<think>
...
</think>

The original final answer is removed during extraction.

🔪 Extraction Process

Original model output:

<think>
Useful reasoning...
</think>

Poor Pashto final answer...

After extraction:

<think>
Useful reasoning...
</think>

The pipeline keeps:

• ✅ User question
• ✅ Model reasoning
• ✅ Empty final answer

And removes:

• ❌ Original final answer

🧪 Why Remove the Final Answer?

Some models can understand a task reasonably well but produce poor Pashto in the final response.

Common problems include:

• grammatical errors
• unnatural wording
• semantic drift
• repeated phrases
• incorrect verbs
• strange translations
• unrelated content

Instead of using a poor final answer as training data, this dataset attempts to preserve the useful reasoning.

🔄 Multi-Model Pipeline

The extracted reasoning can later be passed to another model that is stronger at Pashto generation.

              Teacher Model
                   │
                   ▼
                Question
                   │
                   ▼
              Reasoning 🧠
                   │
                   ▼
        ┌────────────────────┐
        │   Pashto Writer    │
        │       Model        │
        └────────────────────┘
                   │
                   ▼
          Natural Pashto Answer

This allows different models to contribute their strongest abilities.

For example:

Model A → reasoning
Model B → Pashto generation
Model C → judging / filtering

🧹 Quality Control

The extraction pipeline checks basic structural requirements.

A sample should contain:

• a valid JSON object
• a messages field
• a user question
• an assistant message
• a non-empty reasoning section
• a valid <think>...</think> block

Duplicate samples can also be removed.

Accepted samples are written to the cleaned dataset, while rejected samples are stored separately for inspection.

📊 Current Dataset

The initial extraction experiment produced:

• 62 total samples
• 62 accepted samples
• 0 rejected samples
• 100% structural acceptance rate

The 100% acceptance rate refers to structural extraction only. It does not mean that every reasoning sample is linguistically or semantically perfect.

Further quality filtering can be performed in later stages.

🧠 Future Development

Stage 1 — Brain Extraction

Question + Reasoning

Stage 2 — Brain Quality Judge

Evaluate whether the reasoning:

• follows the question
• preserves the main characters
• preserves the event
• proposes a sensible solution
• avoids irrelevant discussion
• is useful to another generation model

Stage 3 — Pashto Writer

Use accepted reasoning to generate the final answer with a stronger Pashto-generation model.

Stage 4 — Final Judge

Evaluate the generated answer for:

• natural Pashto
• grammar
• semantic accuracy
• story coherence
• simple vocabulary
• child suitability

⚠️ Limitations

This is an experimental dataset.

The reasoning was generated by language models and may contain:

• incorrect assumptions
• unnecessary meta-commentary
• repetitive reasoning
• unnatural Pashto
• incorrect terminology
• incomplete planning
• model-specific artifacts

Therefore, the dataset should not be treated as human-authored reasoning.

🤝 Intended Use

This dataset is intended for:

• Pashto LLM research
• reasoning experiments
• teacher-model experiments
• multi-model pipelines
• synthetic data generation
• Pashto instruction tuning
• reasoning-to-generation experiments
• model evaluation
• studying the separation of reasoning and final language generation

🚫 Not Intended As

This dataset is not intended to be:

• a collection of final Pashto answers
• a general-purpose Pashto corpus
• a factual knowledge database
• a replacement for human-written Pashto data

The central purpose is reasoning extraction.

📜 License

This dataset is released under the CC BY 4.0 license.

👤 Author

Created by nassimjp as part of ongoing experiments in Pashto language AI, synthetic data generation, model evaluation, and multi-model training pipelines.

---

🧠 Philosophy

که د ماډل خوله خرابه وي، ټول ماډل مه غورځوه؛ شاید مغز یې لا هم ګټور وي.

Keep the brain. Replace the mouth. 🧠🔪🐍