---
language:
  - en
license: apache-2.0
tags:
  - thermodynamics
  - exergy-analysis
  - second-law
  - industrial-energy
  - domain-specific
  - fine-tuned
  - gguf
  - ollama
  - edge-deployment
base_model: Qwen/Qwen3-8B
model_type: causal-lm
quantized_by: unsloth
pipeline_tag: text-generation
library_name: llama.cpp
datasets:
  - custom
---

# EntropyHunter-8B v0.4 — Exergy Analysis Specialist

**A fine-tuned LLM specialized in second-law thermodynamic (exergy) analysis of industrial equipment.**

EntropyHunter performs rigorous exergy calculations including energy/exergy balances, entropy generation, exergoeconomic analysis, and thermodynamic optimization — tasks that general-purpose LLMs struggle with despite their broad knowledge.

## Key Features

- **Domain Expert:** Trained specifically for exergy (second-law) analysis across 7 equipment types and 6 analysis categories
- **Verified Training Data:** 1,369 examples generated by Claude Opus 4.6 and validated through thermodynamic consistency checks (energy balance, Gouy-Stodola, Bejan number)
- **Edge Deployable:** 8B parameters, Q4_K_M quantization (4.7 GB) — runs on laptops, Jetson Orin, and industrial edge devices without cloud connectivity
- **No Cloud Required:** Industrial data stays on-premise — EU data sovereignty compliant

## Model Details

| Property | Value |
|----------|-------|
| Base Model | Qwen3-8B |
| Fine-tuning | LoRA (r=16, α=32) via Unsloth |
| Training Data | 1,235 train / 134 val examples (ChatML) |
| Teacher Model | Claude Opus 4.6 (knowledge distillation) |
| Quantization | Q4_K_M (4.7 GB) |
| Context Length | 8,192 tokens |
| Training Loss | 0.408 |
| Thinking Mode | Disabled (direct response, no chain-of-thought tokens) |

## Supported Analysis Types

| Analysis | Description |
|----------|-------------|
| **Basic Exergy** | Exergy destruction, exergetic efficiency, irreversibility |
| **Entropy Generation** | Entropy generation mechanisms, Gouy-Stodola theorem |
| **Exergoeconomic** | Cost of exergy destruction, thermoeconomic evaluation |
| **What-If Comparison** | Parameter sensitivity, operating condition optimization |
| **Avoidable/Unavoidable** | Splitting exergy destruction into improvement potential |
| **Hotspot Detection** | Multi-equipment system bottleneck identification |

## Supported Equipment

Compressors, heat exchangers, boilers, steam turbines, pumps, chillers, dryers, and multi-equipment factory systems — across 48 industrial subtypes.

## Quick Start — Ollama

```bash
# Download and create model
ollama create entropy-hunter -f Modelfile

# Run
ollama run entropy-hunter "Perform exergy analysis for a centrifugal compressor with inlet air at 25°C, 101.325 kPa; outlet at 215°C, 800 kPa; mass flow 2.5 kg/s; power input 520 kW; isentropic efficiency 78%. Dead state: T₀ = 25°C, P₀ = 101.325 kPa."
```

### Modelfile

```
FROM ./entropy-hunter-v04-Q4_K_M.gguf

PARAMETER temperature 0.7
PARAMETER top_p 0.8
PARAMETER top_k 20
PARAMETER num_ctx 8192
PARAMETER stop <|im_end|>
PARAMETER stop <|endoftext|>

SYSTEM """You are an expert thermodynamics engineer specializing in exergy (second-law) analysis of industrial equipment. You perform rigorous calculations using dead state conditions T₀ = 25°C (298.15 K), P₀ = 101.325 kPa. /no_think"""
```

## Quick Start — llama.cpp

```bash
./llama-cli \
  --model entropy-hunter-v04-Q4_K_M.gguf \
  --ctx-size 8192 \
  --temp 0.7 \
  --top-p 0.8 \
  --top-k 20 \
  -p "Perform exergy analysis for a shell-and-tube heat exchanger..."
```

## Training Pipeline

```
Claude Opus 4.6 (Teacher)
    │
    ▼
1,500 examples generated (Batch API)
    │
    ▼
Thermodynamic QC Pipeline
├── Energy balance verification (±1%)
├── Gouy-Stodola consistency check
├── Bejan number range (0-1)
├── Exergoeconomic f-factor validation
├── Self-correction aware parsing
└── Balance-line fallback extraction
    │
    ▼
1,369 verified examples (93.4% pass rate)
    │
    ▼
LoRA Fine-tuning (Qwen3-8B, Unsloth)
├── r=16, α=32, lr=1e-4
├── 3 epochs, 465 steps
└── Training loss: 0.408
    │
    ▼
GGUF Export (Q4_K_M, 4.7 GB)
```

## Version History

| Version | Base Model | Data | Score | Notes |
|---------|-----------|------|-------|-------|
| v0.1 | Qwen2.5-7B | 722 examples | 63.5% | First attempt, structural learning |
| v0.2 | Qwen2.5-7B | 885 examples | 85.5% | Stable, "mention vs calculate" fix |
| v0.3 | Qwen2.5-7B | 885 (JSON-free) | 78.3% | Failed experiment, archived |
| **v0.4** | **Qwen3-8B** | **1,369 examples** | **TBD** | **Current — scaffold format, verified distillation** |

### Base Model Comparison (Pre-v0.4)

| Model | Benchmark Score | Notes |
|-------|----------------|-------|
| Base Qwen2.5-7B | 68.4% | No fine-tuning |
| Base Qwen3-8B | 82.6% | No fine-tuning |
| Fine-tuned v0.2 (Qwen2.5-7B) | 85.5% | 885 examples |

## Limitations

- **Arithmetic:** 8B models have inherent arithmetic inconsistency — same inputs can produce slightly different numerical results across runs
- **JSON Output:** Cannot reliably produce structured JSON (known 7B/8B limitation)
- **Language:** English only (optimized for tokenizer efficiency)
- **Scope:** Industrial equipment exergy analysis only — not a general thermodynamics tutor

## Intended Use

- Industrial energy auditors performing second-law analysis
- Engineers identifying thermodynamic inefficiencies in equipment
- Edge deployment for real-time exergy monitoring (IoT integration)
- Educational tool for exergy analysis methodology

## About

Built by [Kemal Düzkar](https://www.linkedin.com/in/kemalduzkar/), chemical engineer and founder of [Olivenet](https://olivenet.io) (Northern Cyprus). EntropyHunter is part of the ExergyLab ecosystem — bringing thermodynamic intelligence to industrial IoT.

**Philosophy:** *Teaching a small model what only large models know — so it can run where large models can't.*

## Citation

```bibtex
@misc{entropyhunter2026,
  title={EntropyHunter-8B: A Domain-Specific LLM for Industrial Exergy Analysis},
  author={Düzkar, Kemal},
  year={2026},
  url={https://huggingface.co/olivenet/entropy-hunter-8b-gguf}
}
```

## License

Apache 2.0