Ejokhan/hydro-expert-llm

🧬 Hydro Expert LLM: Domain-Specific QLoRA Adapter for Hydrology

A QLoRA fine-tuned adapter for Mistral-7B-Instruct-v0.3 trained on 41,958 hydrology Q&A pairs generated from 29,654 scientific papers. Transforms a general-purpose LLM into a hydrology domain expert with 2.8x faster inference and concise, technically accurate answers.

Key Results

Metric	Base Mistral-7B	Fine-tuned
Avg Response Time	18.5s	6.7s
Answer Style	Generic, textbook	Concise, domain-expert
Loss	2.176	1.232 (↓43%)
Token Accuracy	59.7%	70.4%

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

base_model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-7B-Instruct-v0.3",
    quantization_config=bnb_config,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "Ejokhan/hydro-expert-llm")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")

question = "What machine learning models are most effective for streamflow prediction?"
messages = [{"role": "user", "content": question}]
input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=300, temperature=0.3, do_sample=True)
answer = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(answer)

Training Details

Parameter	Value
Base Model	Mistral-7B-Instruct-v0.3
Method	QLoRA (4-bit NF4 + LoRA)
LoRA Rank	16
LoRA Alpha	32
Target Modules	q,k,v,o,gate,up,down projections
Trainable Params	41.9M / 3.8B (1.1%)
Batch Size	16 (4 x 4 gradient accumulation)
Learning Rate	5e-5 (cosine schedule)
Epochs	3
Precision	bf16
Training Time	4h 10m on A100 40GB
Adapter Size	87.6 MB

Training Data

41,958 Q&A pairs generated from 29,654 scientific paper abstracts using Qwen2.5-3B on GPU:

Source	Papers
PubMed	20,292
ArXiv	8,165
Semantic Scholar	1,182
EarthArXiv	62
USGS	15

Topics: streamflow prediction, flood forecasting, water quality monitoring, anomaly detection, hydrological modeling, rainfall-runoff, groundwater, climate science.

Example Comparison

Q: "Compare random forest and gradient boosting for water quality classification"

Base: "Random Forest and Gradient Boosting are two popular machine learning algorithms used for classification problems... Each decision tree in the forest..." (generic textbook)

Fine-tuned: "Random forest performed better than gradient boosting, achieving accuracy of 98.8%." (specific, concise, cites metrics)

Related Projects

1. HydroRAG — RAG system benchmarking 15 retrieval configs over 8,618 papers. Live demo
2. Hydro Expert LLM — This project
3. HydroAgent (planned) — Agentic AI for autonomous hydrological analysis

Infrastructure

Trained on TACC Lonestar6 (NVIDIA A100 40GB) through the NSF NAIRR Pilot program.

Author

Ijaz Ul Haq, Ph.D. — University of Vermont

Google Scholar | GitHub

License

MIT