---
language:
- tr
license: mit
task_categories:
- feature-extraction
size_categories:
- 100K<n<1M
tags:
- embedding-distillation
- student-teacher
pretty_name: Cosmos Corpus Encoded
dataset_info:
  features:
  - name: text
    dtype: string
  - name: teacher_embedding_final
    list: float64
  - name: teacher_embedding_pre_dense
    list: float64
  - name: tabi_input_ids
    list: int64
  - name: cosmos_input_ids
    list: int64
  - name: mursit_input_ids
    list: int64
  - name: mft_input_ids
    list: int64
  splits:
  - name: train
    num_bytes: 6603761113
    num_examples: 224779
  download_size: 1576033082
  dataset_size: 6603761113
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Cosmos Corpus Encoded for Embedding Distillation

This dataset is a pre-tokenized version of [alibayram/cosmos-corpus-0-05-with-embeddings](https://huggingface.co/datasets/alibayram/cosmos-corpus-0-05-with-embeddings), designed for efficient embedding distillation training of MFT and TabiBERT models.

## Dataset Description

- **Source**: `alibayram/cosmos-corpus-0-05-with-embeddings`
- **Language**: Turkish
- **Task**: Embedding Distillation (Teacher-Student Training)
- **Total Examples**: 224,807 (Filtered from 300,000)
- **Max Sequence Length**: 2048 tokens

### Pre-processing & Filtering

The dataset was processed using two different tokenizers to support multiple student architectures:

1. **MFT Tokenizer**: A custom morphologically informed tokenizer.
2. **TabiBERT Tokenizer**: A BERT-based tokenizer with 32k vocabulary.

**Filtering**:

- Original size: 300,000 examples.
- Filtered size: 224,807 examples (~75%).
- Criterion: Both `mft_input_ids` and `tabi_input_ids` must be <= 2048 tokens.
- Sequences longer than 2048 tokens were dropped to ensure efficient training within context limits.

## Dataset Structure

The dataset contains the following columns:

| Column                    | Type          | Description                                                    |
| ------------------------- | ------------- | -------------------------------------------------------------- |
| `text`                    | `string`      | The original raw text content.                                 |
| `mft_input_ids`           | `list[int]`   | Token IDs encoded using the MFT tokenizer.                     |
| `tabi_input_ids`          | `list[int]`   | Token IDs encoded using the TabiBERT tokenizer.                |
| `teacher_embedding_final` | `list[float]` | Final layer embeddings from the teacher model (Gemma-2-9b-it). |

### Data Instances

```python
{
  'text': 'Makine öğrenmesi, verilerden öğrenen algoritmaların çalışılmasıdır.',
  'mft_input_ids': [124, 5921, ...],
  'tabi_input_ids': [101, 2341, ...],
  'teacher_embedding_final': [0.021, -0.054, ...]  # 3584-dimensional vectors
}
```

## Usage

This dataset is optimized for the `EmbeddingDistillationTrainer`. You can load it directly without needing to re-tokenize during training.

```python
from datasets import load_dataset

dataset = load_dataset("alibayram/cosmos-corpus-encoded")
```

### Training Example

To train a model using the `mft_input_ids` column:

```python
from embedding_trainer import EmbeddingDistillationTrainer, EmbeddingTrainerConfig

config = EmbeddingTrainerConfig(
    student_model="alibayram/mft-downstream-task-embeddinggemma",
    input_ids_column="mft_input_ids",  # or "tabi_input_ids"
    embedding_column="teacher_embedding_final",
    loss_type="cosine",
    batch_size=256
)

trainer = EmbeddingDistillationTrainer(config)
trainer.train("alibayram/cosmos-corpus-encoded")
```

## Creation Details

- **Created by**: Ali Bayram
- **Date**: 2026-01-25
- **Teacher Model**: `google/gemma-2-9b-it` (Embeddings extracted via `sartify-llm/Gemma-2-9b-it-v2-embedding`)
- **Processing Script**: `prepare_dataset.py`

## License

MIT