Datasetpublic

pkheria7/indian-legal-opposing-counsel-dataset

⚖️ Indian Legal Opposing Counsel Dataset A combined, preprocessed dataset of 26,326 examples for training an Indian legal opposing counsel AI model. Ready-to-use in ChatML format for SFT training. 📊 Dataset Stats Split Rows Size Train 25,009 65 MB Test 1,317 3.5 MB Total 26,326 69 MB 📦 Sources Source Dataset Rows Content viber1/indian-law-dataset 24,607 Writs, PIL, civil procedure, constitutional law, IPC… See the full description on the dataset page: https://huggingface.co/datasets/pkheria7/indian-legal-opposing-counsel-dataset.

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

0likes67downloads

Dataset Card

⚖️ Indian Legal Opposing Counsel Dataset

A combined, preprocessed dataset of 26,326 examples for training an Indian legal opposing counsel AI model. Ready-to-use in ChatML format for SFT training.

📊 Dataset Stats

Split	Rows	Size
Train	25,009	65 MB
Test	1,317	3.5 MB
Total	26,326	69 MB

📦 Sources

Source Dataset	Rows	Content
viber1/indian-law-dataset	24,607	Writs, PIL, civil procedure, constitutional law, IPC
nisaar/Lawyer_GPT_India	150	Landmark cases, IPC, contract law, constitutional principles
RMani1/indian-legal-dataset-indian-law	1,569	Indian statutes, acts, legal provisions

🗂️ Format

Each row has a messages column in ChatML conversational format (directly compatible with TRL SFTTrainer):

json

{
  "messages": [
    {"role": "system", "content": "You are an experienced opposing counsel specializing in the Indian Constitution..."},
    {"role": "user", "content": "What is the difference between a petition and a plaint in Indian law?"},
    {"role": "assistant", "content": "A petition is a formal request submitted to a court..."}
  ],
  "source": "viber1/indian-law-dataset"
}

⬇️ Download

Option 1: Python (recommended)

python

from datasets import load_dataset
ds = load_dataset("pkheria7/indian-legal-opposing-counsel-dataset")
print(ds)
# DatasetDict({
#     train: Dataset(25009 rows),
#     test: Dataset(1317 rows)
# })

Option 2: Direct JSONL downloads

—📥 train.jsonl (65 MB — 25,009 rows)
—📥 eval.jsonl (3.5 MB — 1,317 rows)
—📥 all.jsonl (69 MB — all 26,326 rows)
—📥 raw_qa_pairs.jsonl (15 MB — just user/assistant, no system prompt)

Option 3: wget / curl

bash

# Full dataset (all splits combined)
wget https://huggingface.co/datasets/pkheria7/indian-legal-opposing-counsel-dataset/resolve/main/data/all.jsonl

# Or just the training split
wget https://huggingface.co/datasets/pkheria7/indian-legal-opposing-counsel-dataset/resolve/main/data/train.jsonl

Option 4: Git clone

bash

git lfs install
git clone https://huggingface.co/datasets/pkheria7/indian-legal-opposing-counsel-dataset

🏋️ Training

Use with the model repo: pkheria7/indian-legal-opposing-counsel

python

from datasets import load_dataset
from trl import SFTConfig, SFTTrainer

ds = load_dataset("pkheria7/indian-legal-opposing-counsel-dataset")

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-7B-Instruct",
    train_dataset=ds["train"],
    eval_dataset=ds["test"],
    args=SFTConfig(
        max_length=2048,
        assistant_only_loss=True,
        push_to_hub=True,
        hub_model_id="your-username/your-model-name",
    ),
)
trainer.train()

📄 License

Apache 2.0