Modelpublic

CaffeineAddict69/qwen7b-water-rural-practices

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

1likes18downloads

Model Card

Qwen 2.5 7B — Rural Water Monitoring Practices Classifier

Fine-tuned version of Qwen/Qwen2.5-7B-Instruct for classifying and extracting good and bad practices in rural water monitoring research papers.

Model Description

This model analyzes text chunks from academic papers and identifies whether they describe good or bad practices for rural water monitoring systems, classifying them across 14 predefined categories (7 good + 7 bad).

The model was fine-tuned with LoRA (r=32) on a dataset of 3516 chunks labeled by Qwen 2.5 72B Instruct AWQ.

Output Format

The model returns structured JSON:

json

{
  "contains_practice": true,
  "practices": [
    {
      "type": "good",
      "categories": ["data_acquisition_technology", "sustainability"],
      "span": "verbatim text from input",
      "explanation": "brief justification",
      "confidence": 0.92
    }
  ],
  "summary": "one-sentence summary"
}

Evaluation Results

Evaluated on a held-out set of 390 chunks:

Metric	Value
JSON parse success	99.7%
Accuracy (contains_practice)	91.0%
Precision	90.2%
Recall	82.8%
F1	0.864
Type agreement (good/bad)	83.8%
Mean categories Jaccard	0.641

Training Details

Parameter	Value
Base model	Qwen/Qwen2.5-7B-Instruct
Method	LoRA (r=32, alpha=64)
Trainable params	80.7M (1.05% of total)
Train samples	3516
Epochs	3
Batch size (effective)	16
Learning rate	2e-4 (cosine schedule)
Final train loss	1.0280
Final eval loss	1.0845
Hardware	NVIDIA A100 SXM4 80GB
Training time	~1 hour

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "CaffeineAddict69/qwen7b-water-rural-practices",
    torch_dtype="bfloat16",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("CaffeineAddict69/qwen7b-water-rural-practices")

system_prompt = '''You are an expert annotator for rural water monitoring research. Analyze text chunks from academic papers and identify practices (good or bad) related to rural water monitoring.

Categories of GOOD practices: data_acquisition_technology, data_management, operation_maintenance, sustainability, community_participation, local_adaptation, scalability.

Categories of BAD practices: inappropriate_technology, non_adaptable_infrastructure, cloud_dependency, high_costs, technical_complexity, centralization, rural_inaccessibility.

Return ONLY valid JSON with this schema:
{"contains_practice": true|false, "practices": [...], "summary": "..."}'''

text = "Your chunk text here..."
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": f"Analyze this chunk:\n\n---\n{text}\n---"},
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=1024, do_sample=False)
response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)

License & Citation

Apache 2.0. Inherits restrictions and rights from the base Qwen 2.5 model.

This model was developed for academic research on appropriate technologies for rural water monitoring.