---
license: eupl-1.2
language:
- en
- nl
- de
- fr
- es
- pt
- it
- zh
- sv
- da
- "no"
- fi
- pl
- cs
- ru
- uk
- el
- hu
- ro
- tr
- ar
tags:
- multilingual
- text-classification
- content-filtering
- multi-dimensional-scoring
- knowledge-distillation
- sustainability
- climate-tech
- clean-energy
- LCSA
- news-analysis
- environmental
library_name: peft
base_model: Qwen/Qwen2.5-1.5B
pipeline_tag: text-classification
metrics:
- mae
- rmse
model-index:
- name: sustainability-technology-v2
  results:
  - task:
      type: text-classification
      name: Multi-dimensional Regression
    metrics:
    - type: mae
      value: 0.717
      name: Test MAE
    - type: mae
      value: 0.654
      name: Validation MAE
---

# Sustainability Technology Filter v2

A fine-tuned **Qwen2.5-1.5B** model with LoRA adapters for multi-dimensional sustainability technology assessment. This model evaluates news articles across **6 LCSA-based dimensions** to identify genuinely impactful sustainability technologies - not just greenwashing or speculation.

## Model Description

### Purpose

This model is designed for automated filtering of sustainability and clean technology news. It scores articles on multiple dimensions derived from the **Life Cycle Sustainability Assessment (LCSA)** framework, enabling:

- **Content curation**: Identify high-quality sustainability technology articles
- **Trend analysis**: Track technology readiness and deployment patterns
- **Research filtering**: Separate substantive innovations from hype

### Key Features

- **Multi-dimensional scoring**: 6 independent LCSA dimensions (0-10 scale)
- **Explicit scope boundaries**: Trained to reject AI/ML papers, consumer electronics, programming tutorials
- **Multilingual support**: 21 languages including EN, DE, FR, ES, PT, NL, ZH, and more
- **Evidence-based**: Focuses on documented deployments and metrics, not announcements

### What's New in v2

- **Improved scope filtering**: Explicit exclusions for off-topic content (AI/ML infrastructure, consumer electronics, military tech)
- **Better dimension independence**: Max correlation 0.61 (down from 0.80+ in v1)
- **Multilingual prefilter**: Keywords in 21 languages for global coverage
- **Lower MAE**: 0.654 validation MAE (vs 0.712 in v1)

---

## Dimensions

The model scores articles on 6 dimensions from the LCSA framework:

### Technology Assessment

| Dimension | Weight | Range | Question |
|-----------|--------|-------|----------|
| **Technology Readiness Level** | 15% | 0-9 | Lab concept to Commercial deployment? |
| **Technical Performance** | 15% | 0-10 | Proven efficiency, reliability, scalability? |
| **Economic Competitiveness** | 20% | 0-10 | Cost-competitive with incumbents? |

### Sustainability Impact

| Dimension | Weight | Range | Question |
|-----------|--------|-------|----------|
| **Life Cycle Environmental Impact** | 30% | 0-10 | Full lifecycle benefits (not just use phase)? |
| **Social Equity Impact** | 10% | 0-10 | Job creation, accessibility, community benefit? |
| **Governance & Systemic Impact** | 10% | 0-10 | Policy alignment, infrastructure readiness? |

### Dimension Descriptions

#### Technology Readiness Level (TRL)
Based on NASA/DOE TRL scale:
- 0: Out of scope (not technology)
- 1-3: Lab/proof of concept
- 4-5: Validated in relevant environment
- 6-7: Demonstrated in operational environment
- 8-9: Commercial deployment at scale

#### Technical Performance
Measures real-world metrics: efficiency improvements, reliability data, scalability evidence, real-world performance.

#### Economic Competitiveness
Life Cycle Cost (LCC) assessment: CAPEX/OPEX competitiveness, learning curve trajectory, market adoption, subsidy dependence.

#### Life Cycle Environmental Impact
Holistic environmental assessment: cradle-to-grave emissions, resource extraction impacts, manufacturing footprint, end-of-life recyclability.

#### Social Equity Impact
Human-centered sustainability: job creation, geographic accessibility, affordability, community acceptance, just transition.

#### Governance & Systemic Impact
System-level readiness: policy alignment, infrastructure compatibility, supply chain maturity, standards and certification.

---

## Performance

### Overall Metrics

| Metric | Validation | Test |
|--------|------------|------|
| **MAE** | 0.654 | 0.717 |
| **RMSE** | 1.14 | 1.22 |

### Per-Dimension Performance (Test Set)

| Dimension | MAE | RMSE |
|-----------|-----|------|
| social_equity_impact | 0.63 | 1.08 |
| economic_competitiveness | 0.67 | 1.15 |
| life_cycle_environmental_impact | 0.69 | 1.10 |
| governance_systemic_impact | 0.77 | 1.28 |
| technical_performance | 0.77 | 1.30 |
| technology_readiness_level | 0.78 | 1.38 |

### Comparison with v1

| Metric | v1 | v2 | Change |
|--------|----|----|--------|
| Validation MAE | 0.712 | **0.654** | -8.1% |
| Test MAE | 0.690 | 0.717 | +3.9% |
| Max dimension correlation | 0.80 | **0.61** | Better independence |

---

## Gatekeeper Rules

### TRL Gatekeeper

**If technology_readiness_level < 3.0 then overall weighted average capped at 2.9**

Rationale: Lab-only technologies cannot achieve high sustainability scores regardless of theoretical potential.

### Tier Classification

| Tier | Weighted Average | Description |
|------|-----------------|-------------|
| **High** | >= 6.0 | Commercial deployment, proven sustainability |
| **Medium** | >= 4.0 | Pilot/early commercial, promising sustainability |
| **Low** | < 4.0 | Lab stage or poor sustainability profile |

---

## Scope Exclusions

The model scores **0 on all dimensions** for off-topic content:

### Excluded Categories

1. **AI/ML Infrastructure** - Model architectures, LLMs, benchmarks (without sustainability application)
2. **Consumer Electronics** - Smartphone reviews, gaming hardware, GPUs
3. **Programming** - Tutorials, frameworks, developer tools
4. **Other** - Military tech, travel, crypto speculation, entertainment

### In-Scope Topics

- Renewable energy (solar, wind, hydro, geothermal, nuclear)
- Electric vehicles and sustainable transport
- Energy storage (batteries, hydrogen, grid storage)
- Carbon capture and emissions reduction
- Circular economy and waste reduction
- Green building and energy efficiency
- Sustainable agriculture and food tech
- AI/ML **applied to** sustainability

---

## Training Details

| Parameter | Value |
|-----------|-------|
| Base Model | Qwen/Qwen2.5-1.5B |
| Training Mode | Knowledge Distillation |
| Oracle Model | Gemini Flash 2.0 |
| Trainable Parameters | 18.5M (1.18% LoRA) |
| Epochs | 3 |
| Batch Size | 8 |
| Learning Rate | 2e-5 |
| Max Length | 512 tokens |

### Data Split

| Split | Examples |
|-------|----------|
| Training | 4,358 |
| Validation | 547 |
| Test | 543 |

---

## Usage

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel
import torch
import numpy as np

# Load model
base_model = AutoModelForSequenceClassification.from_pretrained(
    "Qwen/Qwen2.5-1.5B",
    num_labels=6,
    problem_type="regression"
)
model = PeftModel.from_pretrained(base_model, "jeergrvgreg/sustainability-technology-v2")
tokenizer = AutoTokenizer.from_pretrained("jeergrvgreg/sustainability-technology-v2")

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
    model.config.pad_token_id = tokenizer.pad_token_id

model.eval()

# Score an article
article = "Title: Solar Panel Achieves 30% Efficiency\n\nResearchers developed..."
inputs = tokenizer(article, return_tensors="pt", max_length=512, truncation=True, padding=True)

with torch.no_grad():
    scores = model(**inputs).logits[0].numpy()

dimensions = ["technology_readiness_level", "technical_performance",
              "economic_competitiveness", "life_cycle_environmental_impact",
              "social_equity_impact", "governance_systemic_impact"]
weights = [0.15, 0.15, 0.20, 0.30, 0.10, 0.10]

for dim, score in zip(dimensions, scores):
    print(f"{dim}: {score:.1f}")

weighted_avg = np.average(scores, weights=weights)
if scores[0] < 3.0:
    weighted_avg = min(weighted_avg, 2.9)
print(f"Weighted Average: {weighted_avg:.2f}")
```

---

## Limitations

- **Language**: Training predominantly English; prefilter supports 21 languages
- **High-Tier Data**: Only 0.4% high-tier examples in training
- **Precision**: MAE ~0.7 sufficient for tier classification, not precise scoring
- **Context**: 512 token limit may truncate long articles
- **Temporal**: Trained on 2025-2026 news

---

## Intended Use

### Primary Use Cases
- News aggregation and filtering
- Research monitoring for clean tech
- Content curation for sustainability dashboards
- Trend analysis across sectors

### Out-of-Scope
- Investment decisions (scores content quality, not viability)
- Policy recommendations (requires expert interpretation)
- Academic paper assessment
- Real-time trading

---

## Technical Specifications

- **Architecture**: Qwen/Qwen2.5-1.5B + LoRA (r=8, alpha=16)
- **GPU VRAM**: 4GB minimum, 8GB recommended
- **Inference**: ~30ms/article on RTX 3060

---

## Environmental Impact

- **Hardware**: NVIDIA RTX 4080
- **Training Time**: ~1 hour
- **Carbon**: < 0.1 kg CO2eq

---

## Citation

```bibtex
@misc{sustainability_technology_v2,
  title={Sustainability Technology Filter v2},
  author={NexusMind},
  year={2026},
  url={https://huggingface.co/jeergrvgreg/sustainability-technology-v2}
}
```

---

## Version History

| Version | Date | Changes |
|---------|------|---------|
| v2.0 | 2026-01-14 | Scope exclusions, multilingual prefilter, improved independence |
| v1.0 | 2025-11-27 | Initial LCSA-based model |

---

**Framework Versions**: PEFT 0.17.1, Transformers 4.47+, PyTorch 2.0+