Phipper/pe-energy-infrastructure-training-data

---
language: en
license: apache-2.0
task_categories:
- text-generation
- conversational
tags:
- private-equity
- energy-infrastructure
- finance
- dpo
- sft
- opus-distilled
pretty_name: PE Energy Infrastructure Training Data
size_categories:
- 1K<n<10K
---

# PE Energy Infrastructure Training Data

A training dataset for fine-tuning language models on private equity and energy infrastructure tasks. Created for the [Thurin](https://thurin.ai) project — a PE-focused AI assistant.

## Dataset Summary

| Split | Examples | Format | Description |
|-------|----------|--------|-------------|
| `dpo_pairs.jsonl` | 297 | prompt/chosen/rejected | DPO preference pairs across 5 PE categories |
| `sft_conversations.jsonl` | 804 | messages (chat format) | Supervised fine-tuning conversations |
| `opus_reasoning.jsonl` | 2,308 | messages (chat format) | Claude Opus-distilled reasoning traces |
| **Total** | **3,409** | | |

## Categories

The DPO pairs and SFT conversations cover five core PE domains:

- **Deal Analysis** — LBO screening, due diligence, deal structuring, portfolio company evaluation
- **Financial Modeling** — IRR/MOIC calculations, waterfall distributions, debt sizing, sensitivity analysis
- **Multi-Step** — Complex multi-part PE workflows requiring sequential reasoning
- **Regulatory** — FERC, state PUC, IRA/ITC/PTC, environmental permitting, rate cases
- **Strategy** — Fund strategy, sector allocation, portfolio construction, GP/LP dynamics

## Data Sources

- **Hand-crafted** (127 DPO pairs): Written by a PE professional with domain expertise in energy infrastructure investing
- **Pipeline-generated** (170 DPO + 540 SFT): Generated via an automated pipeline using Claude Opus, with human audit/quality gate
- **Opus-distilled** (2,308 SFT): Reasoning traces distilled from Claude Opus on PE and quantitative problems

## Format

### DPO Pairs
```json
{
"prompt": "We're looking at a 200MW BESS project in ERCOT...",
"chosen": "Detailed, structured PE analysis response...",
"rejected": "Generic or lower-quality response...",
"category": "deal-analysis",
"source": "hand-crafted"
}
```

### SFT Conversations
```json
{
"messages": [
{"role": "user", "content": "Screen this midstream acquisition..."},
{"role": "assistant", "content": "Structured PE analysis..."}
],
"source": "curated-sft"
}
```

## Usage

Designed for:
- **DPO/ORPO training** on PE-specific preference data
- **SFT** for energy infrastructure domain knowledge
- **Evaluation** of model PE capabilities

Tested with MLX (`mlx-lm-lora`) for Apple Silicon fine-tuning. Compatible with any standard training framework (Hugging Face TRL, Axolotl, etc.).

## Training Results

Models fine-tuned on this data:
- [Phipper/qwen3.5-27b-opus-distilled-proposal-c](https://huggingface.co/Phipper/qwen3.5-27b-opus-distilled-proposal-c) — 4.45-4.47 quality score
- Thurin 80B (SFT-v1) — 4.43 PE eval score

## Limitations

- All scenarios are **synthetic** — no real deal data, no proprietary information
- Energy infrastructure focused — may not generalize well to other PE verticals (healthcare, tech, consumer)
- Reflects US regulatory environment (FERC, state PUCs, IRA)
- DPO "rejected" responses are intentionally degraded but still coherent — not adversarial

## Credits

- Created by Nate Baranski ([@Phipper](https://huggingface.co/Phipper))
- Pipeline & evaluation: Bessemer (Claude Code agent)
- Reasoning traces: Claude Opus