---
dataset_info:
  features:
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 77039
    num_examples: 51
  download_size: 33828
  dataset_size: 77039
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
task_categories:
- text-generation
language:
- en
tags:
- legal
- mock-trial
- experimental
- prompt-engineering
---

# Mock-Trial-Data-v2 (Experimental)

This is the **second-generation** training set for the Mock Trial AI project. It represents an intensive experimental phase focused on testing various prompt architectures and instruction-following capabilities for legal simulations.

---

### 🧪 The Research Phase
While **v3** is the current production standard, **v2** was the laboratory where the core logic was forged. This dataset contains the 51 initial instruction-response pairs used to identify the optimal "Loss Floor" for Llama 3.1 8B fine-tuning.

👉 [**Looking for the latest? Check out v3 Data here**](https://huggingface.co/datasets/hobbesthecomputerscientist/mock-trial-datav3)

---

## Dataset Description

- **Curated by:** HobbesTheComputerScientist
- **Focus:** Prompt Engineering & Instruction Adherence
- **Format:** Instruction-Response pairs (JSONL)

## Experimental Scope
In this version, I heavily experimented with **Prompt Permutations** to see which linguistic structures most effectively reduced model hallucination during witness cross-examination. Key experiments included:

* **Constraint Density:** Testing how many rules of evidence could be packed into a single instruction before reasoning degraded.
* **Persona Persistence:** Testing variations in deponent background descriptions to prevent the model from "breaking character."
* **Negative Constraint Testing:** Training the model specifically on what *not* to say (e.g., avoiding legal conclusions).

## Lessons Learned (The Bridge to V3)
The 51 examples in this set revealed that:
1.  Over-constraining the prompt increased the loss by making the model too "robotic."
2.  The model needed more specific "Objection/Response" pairs to handle courtroom interruptions.
3.  These insights led directly to the **Instruction Distillation** methodology used in the current V3 Production Space.

## Licensing
Distributed under the **MIT License**. For educational use in mock trial and legal-tech research.