File size: 4,972 Bytes
dd9ca0d 1128128 344a71c 1128128 344a71c 1128128 344a71c 1128128 344a71c fa38c36 dd9ca0d 5870166 1128128 5870166 1128128 704fd13 1128128 704fd13 1128128 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 | ---
license: mit
task_categories:
- question-answering
- text-generation
language:
- en
tags:
- small-language-models
- slm
- fine-tuning
- short-qa
- low-resource
- hallucination-reduction
pretty_name: Short QA Fine-Tuning Dataset for Small Language Models
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
dataset_info:
features:
- name: prompt
dtype: string
- name: answer
dtype: string
splits:
- name: train
num_bytes: 572797
num_examples: 5020
download_size: 310328
dataset_size: 572797
---
> [!WARNING]
> **This is v2 Of FineTuning dataset so this dataset is update version of old one and this dataset is amazing and factual dataset**
# Dataset Card for Nebulixlabs/FineTuning
## Dataset Summary
**Nebulixlabs/FineTuning** is a short question-answer (QA) pair dataset purpose-built for fine-tuning **very small language models (SLMs)** — roughly in the **1M to 150M parameter** range. Each example consists of a short question (typically **5–6 words**) and a short answer (typically **3–4 words**).
The dataset was created around an empirical observation: very small base models (1M–50M params) can only reliably predict short sequences of about **12–15 words** before accuracy degrades and hallucination increases, regardless of how large their context window is. Training such models on long-form QA pairs pushes them past this reliable generation length, causing hallucinated or incoherent answers.
By constraining both inputs and outputs to short, low-complexity spans, this dataset lets small models learn to answer questions **within their actual reliable generation capacity**, significantly reducing hallucination compared to training on longer QA pairs.
## Supported Tasks
- **Question Answering (short-form):** Train/fine-tune a model to map a short question to a short, factual or direct answer.
- **Hallucination mitigation research:** Useful for studying how output length and training data length affect hallucination rates in small language models.
## Intended Models
This dataset is primarily intended for:
- Small language models (SLMs) in the **1M–50M parameter** range (primary use case).
- Mid-size models up to **~150M parameters**, since empirical experiments showed models around this size are also prone to high hallucination rates and benefit from short-pair fine-tuning.
It is **not recommended** for larger/mature models (e.g., 1B+ parameters), which can already handle longer context and generation reliably and may be under-utilized by such short pairs.
## Languages
English (`en`).
## Dataset Structure
### Data Instances
Each instance is a short question-answer pair. Example:
```json
{
"prompt": "What is the capital of France?",
"answer": "Paris."
}
```
### Data Fields
| Field | Type | Description |
|------------|--------|---------------------------------------------------|
| `prompt` | string | A short question, typically 5–6 words long. |
| `answer` | string | A short, direct answer, typically 3–4 words long. |
### Data Splits
| Split | # Examples |
|-------|-----------|
| train | 5k |
## Dataset Creation
### Rationale
Experiments showed that very small base models (1M–50M parameters) reliably generate correct output only up to **12–15 words**, independent of context length. Fine-tuning such models on longer QA pairs causes hallucination even within a large context window. This dataset keeps both question and answer short so the generation task stays within the model's reliable output range, which:
- Reduces hallucination during inference.
- Makes training signal cleaner and easier for tiny models to fit.
- Allows models as large as ~150M parameters (which also showed elevated hallucination rates in testing) to benefit from the same short-pair format.
### Curation Rationale / Source Data
> *This dataset is synthatic generated by Sarwam 105B Conversation model*
### Annotations
> *This dataset is tested and verified*
## Considerations for Using the Data
### Limitations
- Short QA format is **not suitable** for tasks requiring long-form reasoning, multi-turn dialogue, or detailed explanations.
- Intended specifically for small-parameter models; larger models may not show meaningful benefit.
- Answer brevity may limit factual nuance (e.g., no room for caveats or partial answers).
## Additional Information
### Licensing
Specify the license under which this dataset is released (e.g., MIT, CC-BY-4.0, Apache-2.0).
### Citation
```bibtex
@dataset{nebulixlabs_finetuning,
title = {Nebulixlabs/FineTuning: Short QA Pairs for Small Language Model Fine-Tuning},
author = {Nebulixlabs},
year = {2026},
url = {https://huggingface.co/datasets/Nebulixlabs/FineTuning}
}
```
### Contact
For questions, issues, or contributions, open a discussion on the dataset's Hugging Face repository page. |