---
pretty_name: FEJI First Fine-Tune
language:
  - tr
tags:
  - audio
  - music
  - turkish
  - turkish-folk
  - ace-step
  - fine-tuning
task_categories:
  - text-to-audio
size_categories:
  - n<1K
license: other
---

# FEJI First Fine-Tune

FEJI First Fine-Tune is a Turkish folk music dataset prepared for ACE-Step 1.5
LoRA fine-tuning experiments. It contains 201 audio examples with aligned
ACE-Step metadata for caption-conditioned music generation.

The dataset was exported from the local `finetune-dataset/` folder and uploaded
as Parquet shards with embedded audio. Each row contains one audio sample plus
metadata fields used by the ACE-Step training and dataset-builder workflow.

## Dataset Details

- **Rows:** 201
- **Default split:** `train`
- **Primary language:** Turkish (`tr`)
- **Primary domain:** Turkish folk / Turku-style music
- **ACE-Step custom tag:** `fejiturkishmakam`
- **Audio format in source folder:** WAV
- **Published format:** Hugging Face Parquet with embedded `audio` column

## Intended Use

This dataset is intended for:

- ACE-Step 1.5 LoRA / adapter fine-tuning.
- Experiments in Turkish folk music generation.
- Caption-conditioned music generation research using makam, usul, region,
  instrumentation, lyrics, and tempo metadata.
- Reproducible dataset loading through Hugging Face `datasets`.

## Columns

Important columns include:

| Column | Description |
| --- | --- |
| `audio` | Embedded audio sample loaded through `datasets.Audio`. |
| `id` | Stable sample identifier, matching the local audio filename stem. |
| `audio_path` | Original local relative audio path from the ACE-Step dataset JSON. |
| `filename` | Original audio filename. |
| `caption` | Main ACE-Step text-conditioning caption. |
| `lyrics` | Structured lyrics used by the lyrics branch. |
| `raw_lyrics` | Unformatted lyrics text. |
| `formatted_lyrics` | Lyrics with section markers such as `[Verse]` / `[Chorus]`. |
| `bpm` | Estimated or assigned tempo. |
| `keyscale` | Key / scale metadata. |
| `timesignature` | Time signature metadata. |
| `duration` | Duration in seconds. |
| `language` | Language code. |
| `is_instrumental` | Whether the sample is instrumental. |
| `custom_tag` | Trigger tag used for adapter training. |
| `makam` | Turkish makam label when available. |
| `usul` | Rhythmic/usul label when available. |
| `region` | Regional label when available. |
| `instruments` | Instrument list. |
| `vocal` | Vocal-performance descriptor. |
| `source_url` | Source URL recorded in the local manifest. |
| `youtube_id` | YouTube ID recorded in the local manifest. |
| `playlist` | Source playlist ID recorded in the local manifest. |

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("alibayram/feji-first-finetune", split="train")
sample = dataset[0]

print(sample["caption"])
print(sample["lyrics"])
print(sample["audio"])
```

The `audio` field is decoded by Hugging Face `datasets` as an audio dictionary
containing the waveform array, sampling rate, and original path information.

## ACE-Step Usage Notes

For ACE-Step fine-tuning, use the metadata as follows:

- Put instrument, makam, region, style, and performance cues in `caption`.
- Use `lyrics` / `formatted_lyrics` for the lyrics branch.
- Keep the trigger tag `fejiturkishmakam` consistent across training and
  generation prompts.
- Use the BPM, key, and time-signature columns when building or validating
  structured training samples.

The local export script used for this dataset validates that every JSON sample
has a matching audio file before upload and preserves every sample metadata
field from `dataset.json`.

## Limitations and Rights

This dataset card documents the technical dataset structure and intended model
training workflow. Source-rights, redistribution rights, and downstream
commercial-use permissions should be verified before public or commercial use.