---
dataset_info:
  features:
  - name: messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: source
    dtype: string
  - name: topic
    dtype: string
  splits:
  - name: train
    num_bytes: 73571533
    num_examples: 71037
  - name: eval
    num_bytes: 8174614
    num_examples: 7893
  download_size: 74336872
  dataset_size: 81746147
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: eval
    path: data/eval-*
task_categories:
- question-answering
- reinforcement-learning
- text-generation
size_categories:
- 10K<n<100K
tags:
- healthcare
- first-aid
- emergency-response
- instruction-tuning
- supervised-finetuning
- rag
- medical-qa
- grounded-generation
- safety
- domain-specific-llm
---


# FirstAid-Treatment IT

## Dataset Description

**FirstAid-Treatment IT** is an instruction tuning and question-answering dataset for training domain-specialized language models on **first aid, emergency response, procedural medical support, and safety guidance**.

The dataset consists of grounded conversational training examples built from source-aligned first-aid reference material and emergency response documentation.

It is designed for:

- Supervised Fine-Tuning (SFT)
- Question Answering (QA)
- Retrieval-Augmented Generation (RAG)
- Grounded medical assistant training
- Safety-focused instruction-following models

## Supported Tasks

This dataset supports several NLP tasks:

### 1. Instruction Fine-Tuning
Examples are formatted as chat-style instruction data:

- Summarize this passage
- Extract procedural steps
- List key takeaways
- Answer grounded domain questions

Suitable for fine-tuning:
- :contentReference[oaicite:2]{index=2} Llama models
- :contentReference[oaicite:3]{index=3} models
- :contentReference[oaicite:4]{index=4} Qwen models
- Medical domain assistants

---

### 2. Question Answering

Supports:
- Extractive QA
- Grounded QA
- Procedural QA
- Multi-step reasoning QA

Example tasks:

Question:
What should be prioritized in massive hemorrhage response?

Answer:
Follow immediate bleeding control protocols, prioritize life-threatening hemorrhage intervention, and escalate emergency care.

---

### 3. Retrieval-Augmented Generation

Useful for:
- Grounded response generation
- Hallucination reduction
- Citation-aware medical assistants
- Retrieval evaluation benchmarks

---

# Dataset Summary

- Total rows: 78,930
- Train split: 71K+
- Format: Parquet
- Modality: Text
- Message-format conversational samples
- Source-grounded instruction-response data

## Topics Covered

Includes diverse first aid topics such as:

- Trauma management
- Bleeding control
- Poisoning response
- Respiratory emergencies
- Pediatric emergencies
- Disaster response
- Toxic exposure
- Tactical casualty care
- Infection control
- Crowd crush and stampede support
- Hazardous materials response
- Radiation exposure first aid
- Decompression and dive injuries
- Neonatal emergencies
- Mental health crisis support

---

# Dataset Structure

Each example follows chat-style instruction tuning format:

```python
{
 "messages": [
   {
      "role":"system",
      "content":"You are a domain expert assistant. Answer only from provided context."
   },
   {
      "role":"user",
      "content":"Extract procedural steps..."
   },
   {
      "role":"assistant",
      "content":"1. Assess scene safety..."
   }
 ],
 "source":"firstaid_114_massive_hemorrhage_march.pdf_chunk_449",
 "topic":"massive hemorrhage management"
}
```

## Features

| Feature | Type |
|--------|------|
| messages | list |
| source | string |
| topic | string |

---

# Dataset Creation

## Source Data

Built from structured first aid and emergency care reference materials, guidelines, and procedural source documents.

Examples are generated from source chunks using grounded instruction synthesis patterns:

- summarization prompts
- procedural extraction prompts
- knowledge QA prompts
- key takeaway generation

## Data Generation Pipeline

High-level pipeline:

1. Source ingestion  
2. Document chunking  
3. Topic extraction  
4. Instruction synthesis  
5. Grounded answer generation  
6. Quality filtering  
7. Conversational formatting

---

# Intended Uses

Recommended:

✅ SFT for medical assistants  
✅ Domain adaptation  
✅ Grounded QA systems  
✅ RAG benchmarking  
✅ Safety-aligned assistant training  
✅ Medical instruction following research

---

# Out-of-Scope Uses

Not intended for:

- Real-time emergency diagnosis
- Clinical decision making
- Autonomous medical advice
- Replacement for licensed professionals
- Unsupervised deployment in high-risk settings

This is a training dataset, not medical advice.

---

# Bias, Risks, and Limitations

## Limitations

- Coverage limited to included sources
- May not reflect jurisdiction-specific protocols
- Not a substitute for updated clinical guidance
- Synthetic instruction generation may introduce formatting bias

## Safety Risks

Models trained solely on this data may:
- Overgeneralize procedures
- Omit escalation contexts
- Produce incomplete emergency recommendations

Human oversight is strongly recommended.

---

# Fine-Tuning Example

```python
from datasets import load_dataset

ds = load_dataset(
    "nuhmanpk/FirstAid-Treatment_it"
)

print(ds["train"][0]["messages"])
```

Example SFT formatting:

```python
tokenizer.apply_chat_template(
  sample["messages"],
  tokenize=False
)
```

Works well for chat-style SFT pipelines using:
- TRL SFTTrainer
- Axolotl
- Unsloth
- Transformers Trainer

---

# Benchmark / Evaluation Ideas

Can be evaluated on:

- Exact Match (EM)
- F1 for QA
- Faithfulness
- Groundedness
- Hallucination rate
- Instruction-following accuracy

Potential benchmark tasks:
- Procedural step extraction
- Emergency response QA
- Safety reasoning
- Medical RAG evaluation

---

# Citation

```bibtex
@dataset{nuhmanpk_firstaid_treatment_it,
 author = {Nuhman PK},
 title = {FirstAid-Treatment IT},
 year = {2026},
 publisher = {Hugging Face},
 url = {https://huggingface.co/datasets/nuhmanpk/FirstAid-Treatment_it}
}
```

---

# Creator

Created by **Nuhman PK**

## Links

- GitHub: https://github.com/nuhmanpk
- Hugging Face: https://huggingface.co/nuhmanpk
- Kaggle: https://www.kaggle.com/nuhmanpk
- LinkedIn: https://linkedin.com/in/nuhmanpk
- Portfolio: https://nuhmanpk.github.io/portfolio

---

If you use this dataset in research or fine-tuning, cite the dataset and share your results.