---
language:
- as
- en
license: apache-2.0
size_categories:
- 10K<n<100K
task_categories:
- text-generation
- question-answering
- conversational
tags:
- assamese
- indic
- instruction-tuning
- sft
- gemma
- alpaca
- ai4bharat
dataset_info:
  features:
  - name: id
    dtype: string
  - name: source
    dtype: string
  - name: system_prompt
    dtype: string
  - name: messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: prompt
    dtype: string
  - name: response
    dtype: string
  - name: text
    dtype: string
  - name: language
    dtype: string
  - name: num_turns
    dtype: int64
  - name: char_count
    dtype: int64
  splits:
  - name: train
    num_bytes: 476367883
    num_examples: 58833
  - name: validation
    num_bytes: 24514134
    num_examples: 3096
  download_size: 500882017
  dataset_size: 500882017
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
---

# Assamese SFT Dataset v1 (`ananddey/assamese-sft-dataset-v1`)

An industry-standard, curated, and deduplicated Supervised Fine-Tuning (SFT) dataset for training Assamese language models and conversational AI assistants.

## 📊 Dataset Summary

* **Total Samples:** 61,929 instruction-response pairs
* **Train Split:** 58,833 samples
* **Validation Split:** 3,096 samples
* **Languages:** Assamese (`as`), English (`en`)
* **Primary Use Case:** SFT / Instruction Fine-Tuning for generative LLMs (e.g. Gemma, Llama, Mistral)

## 🏛️ Sources & Composition

1. **`Telugu-LLM-Labs/assamese_alpaca_yahma_cleaned_filtered`** (~28.7k rows):
   - Single-turn general task instructions, reasoning, translation, and structured task following in Assamese.
2. **`ai4bharat/indic-align` (Dolly_T)** (~15.0k rows):
   - Factual Q&A, summarization, creative writing, and text classification in Assamese.
3. **`ai4bharat/indic-align` (OpenAssistant_T)** (~19.9k rows):
   - Multi-turn conversational flow, technical explanations, and natural dialogue.
4. **Filtering & Deduplication**:
   - Filtered for minimum Assamese script ratio (>20%), stripped empty/corrupted entries, and deduplicated using MD5 content hashing (1,706 duplicates removed).

## 📋 Schema Format

| Field | Type | Description |
|---|---|---|
| `id` | `string` | Unique record identifier (e.g., `alpaca_as_023018`) |
| `source` | `string` | Provenance dataset tag |
| `system_prompt` | `string` | Assistant persona in Assamese (`আপুনি এজন সহায়ক, বিনম্ৰ আৰু সৎ অসমীয়া এআই সহকাৰী।`) |
| `messages` | `list[dict]` | Standard OpenAI/HF Chat template (`[{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]`) |
| `prompt` | `string` | User instruction |
| `response` | `string` | Target completion |
| `text` | `string` | Gemma formatted chat text (`<start_of_turn>user\n...<end_of_turn>\n<start_of_turn>model\n...<end_of_turn>`) |
| `language` | `string` | `"as"` |
| `num_turns` | `int` | Number of conversation turns |
| `char_count` | `int` | Total character count |

## 🚀 Usage

### Loading with Hugging Face `datasets`

```python
from datasets import load_dataset

dataset = load_dataset("ananddey/assamese-sft-dataset-v1")
print(dataset["train"][0])
```

### Loading with TRL `SFTTrainer`

```python
from trl import SFTTrainer
from transformers import TrainingArguments

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"],
    dataset_text_field="text",
    max_seq_length=2048,
    args=TrainingArguments(
        output_dir="./gemma-4-as-sft",
        learning_rate=1e-4,
        num_train_epochs=2,
    )
)
```