Upload README.md with huggingface_hub

README.md

@@ -1,58 +1,97 @@
 ---
-dataset_info:
-- config_name: humour_chat
-  features:
-  - name: messages
-    list:
-    - name: role
-      dtype: string
-    - name: content
-      dtype: string
-  splits:
-  - name: train
-    num_bytes: 2133898
-    num_examples: 5017
-  download_size: 2134780
-  dataset_size: 2133898
-- config_name: math_reasoning
-  features:
-  - name: messages
-    list:
-    - name: role
-      dtype: string
-    - name: content
-      dtype: string
-  splits:
-  - name: train
-    num_bytes: 8603561
-    num_examples: 20504
-  download_size: 8604153
-  dataset_size: 8603561
-- config_name: merged_genz_chat
-  features:
-  - name: messages
-    list:
-    - name: role
-      dtype: string
-    - name: content
-      dtype: string
-  splits:
-  - name: train
-    num_bytes: 1087638
-    num_examples: 4590
-  download_size: 1075799
-  dataset_size: 1087638
-configs:
-- config_name: humour_chat
-  data_files:
-  - split: train
-    path: humour_chat/train-*
-- config_name: math_reasoning
-  data_files:
-  - split: train
-    path: math_reasoning/train-*
-- config_name: merged_genz_chat
-  data_files:
-  - split: train
-    path: merged_genz_chat/train-*
+license: mit
+task_categories:
+- text-generation
+language:
+- en
+tags:
+- sft
+- conversational
+- instruction-tuning
+- multi-genre
+- math-reasoning
+- humour
+- genz
+size_categories:
+- 10K-100K
 ---
+
+# Filthy-data-SFT
+
+This is a highly curated, cleaned, and professionally reformatted multi-genre conversational dataset prepared specifically for **Supervised Fine-Tuning (SFT)** of custom LLMs.
+
+The original repository (`Arko007/Filthy-data`) contained unstructured files with heterogeneous schemas (e.g., nesting fields, different keys, varied turn structures). This version standardizes everything into clean, unified message list-of-dicts structures.
+
+## Subsets Available
+This dataset is divided into three major target configurations:
+1. **`math_reasoning`** (20504 rows): Complex math and step-by-step reasoning sequences.
+2. **`humour_chat`** (5017 rows): Humorous conversational pairs.
+3. **`merged_genz_chat`** (4590 rows): Unified and cleaned dialogue incorporating both normal GenZ slang dialogues and the extreme colloquial dialogue sets.
+
+---
+
+## Dataset Anatomy & Features
+
+Every example in any of the subsets strictly follows the uniform schema:
+- `messages` (list of dicts):
+  - `role` (string): Either `"user"` or `"assistant"`.
+  - `content` (string): Text content of the dialogue turn.
+
+### Sample Data Structure
+```json
+{
+  "messages": [
+    {
+      "role": "user",
+      "content": "Yo, what is the vibe today?"
+    },
+    {
+      "role": "assistant",
+      "content": "No cap, we are just cooling out and vibing!"
+    }
+  ]
+}
+```
+
+---
+
+## Quick Start & Usage
+
+You can load individual subsets directly from the Hugging Face Hub using the `datasets` library:
+
+```python
+from datasets import load_dataset
+
+# Load math reasoning
+math_ds = load_dataset("Arko007/Filthy-data-SFT", "math_reasoning")
+
+# Load merged GenZ chats
+genz_ds = load_dataset("Arko007/Filthy-data-SFT", "merged_genz_chat")
+
+print(genz_ds["train"][0])
+```
+
+### Finetuning Template (SFTTrainer)
+This dataset format is instantly compatible with Hugging Face's `SFTTrainer` and respects models' chat templates natively:
+
+```python
+from trl import SFTTrainer
+from transformers import TrainingArguments
+
+trainer = SFTTrainer(
+    model=model,
+    train_dataset=genz_ds["train"],
+    max_seq_length=1024,
+    dataset_text_field="messages",
+    packing=False,
+    args=TrainingArguments(
+        output_dir="./sft_results",
+        per_device_train_batch_size=4,
+        gradient_accumulation_steps=4,
+        logging_steps=10,
+        save_steps=100,
+        learning_rate=2e-5
+    )
+)
+trainer.train()
+```