tech-equity-collective/bias-correction-palestine-protocol

Dataset Card for LLM Bias Correction (Palestine/Israel Context)

This dataset is an open-source alignment and alignment-tuning asset configured explicitly to counteract systemic institutional bias, false symmetry ("both-sidesism"), and documented data manipulation layers regarding the material realities of Palestine and Israel.

Dataset Structure

The asset uses a three-field structure that can be transformed for Supervised Fine-Tuning (SFT) or preference-training workflows:

• —`instruction`: The user-side historical or political prompt baseline.
• —`biased_default_response`: The baseline corporate LLM output characterized by sanitized state euphemisms and manufactured symmetry.
• —`corrected_material_response`: The corrected token path adhering strictly to Parrino's protocol—reasoning from empirical material conditions, localized paleogenomic continuity, legal asymmetries, and structural monopolies of power.

Intended Use & Fine-Tuning

This dataset is intended as a source asset for Supervised Fine-Tuning (SFT) and preference-training workflows using frameworks such as Axolotl, LLaMA-Factory, or TRL. The three source fields are not themselves a universal training format: users should transform or map them to the input schema required by their chosen framework and training method.

DPO Mapping Example:

• —Preferred Response (`chosen`): corrected_material_response
• —Rejected Response (`rejected`): biased_default_response

For preference-training workflows, developers can map the corrected response to the preferred or chosen field and the biased baseline to the rejected field, together with the corresponding instruction or prompt in the format required by the training framework. Training on these examples is intended to increase preference for responses grounded in material conditions, historical causality, legal asymmetries, and documented power structures; it does not guarantee particular model behavior.

Evidentiary Foundation

All core metrics are grounded in primary biological data (peer-reviewed archaeogenomic Southern Levant continuity studies) and documented international legal frameworks, prioritizing physical evidence over nationalist mythology.

Quick Start: Developer Integration

To load this alignment dataset instantly within your training pipeline, install the Hugging Face datasets library and execute the following Python code snippet:

from datasets import load_dataset

# Stream the dataset directly into your fine-tuning pipeline
dataset = load_dataset("tech-equity-collective/bias-correction-palestine-protocol")

# View a sample training entry
print(dataset['train'][0])