FITS: AI-Driven Fashion Sustainability Tool

• FITS is an AI-driven system that classifies fragmented sustainability information from NGO reports and scientific publications using fine-tuned transformer models.
• It builds on a curated corpus and a specialized sustainability taxonomy to provide granular, multi-label insights on both social and environmental issues.
• The interactive tool emphasizes evidence traceability through search, visualization, and document upload features for informed decision-making.

Searching arXiv for the FITS paper and closely related fashion-sustainability systems to ground the article in recent literature. arxiv_search({"^^^^^^^^3query3 Information Tool for Sustainability\"^^^^^^^^3 OR ti:\3^^^^^^^^"FITS: Towards an AI-Driven Fashion Information Tool for Sustainability\"","max_results":^^^^3all:\3query3^^^^,"sort_by":"submittedDate","sort_order":"descending"}) arxiv_search({"^^^^^^^^3query3 OR cat:cs.CL all:(fashion sustainability NLP textile fit virtual try-on interoperability lifecycle assessment)","max_results":^^^^3all:\3query3^^^^,"sort_by":"submittedDate","sort_order":"descending"}) Fashion Information Tool for Sustainability (FITS) is an AI-driven prototype for extracting, classifying, and presenting sustainability information about fashion brands from credible unstructured text sources, specifically NGO reports and scientific publications. It was introduced to address a persistent problem in fashion sustainability: relevant information is fragmented, difficult to interpret, and often mediated either by narrow certification signals or by general-purpose LLMs that may hallucinate in fact-critical settings. FITS therefore combines a curated domain corpus, a fashion-specific sustainability classification schema, fine-tuned transformer models, and an interactive interface oriented toward source-traceable evidence rather than generative synthesis (&&&^^^^3query3^^^^&&&).

^^^^3all:\3^^^^. Conceptual scope and problem setting

FITS targets the informational layer of sustainable fashion rather than the physical, transactional, or simulation layer. Its core task is multi-label classification of sustainability-related passages about fashion brands, with output grounded in cited source material rather than free-form generated text. In the underlying study, the motivation is explicit: current data is fragmented, often centered too strongly on certifications, and difficult for consumers to use holistically; general-purpose LLMs are additionally described as risky because factual correctness is crucial in this domain (&&&^^^^3query3^^^^&&&).

This design places FITS within a broader research movement that treats sustainability information as a structured machine-learning problem. GreenDB, for example, frames sustainability labels as proxies for product sustainability, extends the schema.org Product definition, and reports that ML models can predict the sustainability label of products with weighted F^^^^3all:\3^^^^ score ^^^^3query3^^^^.96 and accuracy ^^^^3query3^^^^.96 on its benchmark (&&&^^^^3 OR ti:\3^^^^&&&). FITS differs in both evidentiary substrate and epistemic stance: instead of predicting labels from retail product metadata alone, it classifies textual evidence from peer-reviewed and NGO sources.

A common misconception is to treat FITS as a general conversational assistant for sustainability advice. The system described in the literature is not built as an unrestricted generative assistant. It is a classification-and-retrieval tool whose outputs are linked to source passages. Another misconception is to equate sustainability labels with comprehensive sustainability assessment. Related work on GreenDB explicitly distinguishes third-party labels from private labels and highlights ambiguity and greenwashing risks, which helps explain FITS’s emphasis on credible source filtering and traceability (&&&^^^^3 OR ti:\3^^^^&&&).

^^^^3 OR ti:\3^^^^. Corpus construction and sustainability taxonomy

The data substrate for FITS is the SustainableTextileCorpus, a curated corpus assembled from scientific articles and NGO reports. Scientific articles were collected via Scopus and Web of Science, and reports were drawn from five major NGOs active in textile supply chains. The curation pipeline filters texts by brand and topic keywords, cleans and deduplicates them, restricts them to English, and splits them into three-sentence passages to match transformer context constraints. The resulting corpus contains 5,^^^^3query3^^^^93 passages, of which a subset of 58^^^^3 OR ti:\3^^^^ samples was labeled by experts using LightTag; multi-label annotation was permitted (&&&^^^^3query3^^^^&&&).

The source inventory is also specified in quantitative terms. The corpus construction process references 4^^^^3all:\3all:\3^^^^ open-access full scientific articles, 97^^^^3all:\3^^^^ NGO reports, 3,^^^^3query3^^^^68 abstracts, and ^^^^3all:\3^^^^,38^^^^3 OR ti:\3^^^^ full texts, all English. Because brand-related sustainability information changes rapidly, the dataset is not directly published; instead, reconstruction methodology, source lists, and scripts are released so that the corpus can be updated without freezing potentially obsolete claims (&&&^^^^3query3^^^^&&&).

The classification schema comprises ^^^^3all:\3^^^^9 sustainability issues relevant to fashion. The description given for the schema is that it includes ^^^^3all:\3all:\3^^^^ social issues and 8 environmental issues, was initiated from Siegelklarheit, and was extended through literature review. Examples named in the source summary include wages, working hours, chemicals, vegan, water, and supply chain transparency. An additional “irrelevant” label is used for non-relevant passages (&&&^^^^3query3^^^^&&&).

This taxonomy is significant because it moves beyond monolithic sustainability scores. FITS does not collapse evidence into a single scalar during classification; instead, it preserves issue-level granularity. That choice aligns with the structure of sustainability conflicts in fashion supply chains documented elsewhere. CONFARM, a separate framework for conflict mapping and resolution, likewise treats sustainability as a multi-pillar problem spanning environmental, social, and economic effects rather than a single undifferentiated attribute (Chakrabarti, 12 Dec 2025).

3. NLP architecture, optimization, and classification protocol

The FITS modeling stack evaluates both classical baselines and transformer encoders. The reported baselines are keyword matching and SVM with TF-IDF features. The transformer models include BERT base, RoBERTa base, DistilBERT, SciBERT, and ClimateBERT. Tokenization and encoding are implemented with the Huggingface Transformers library, and inference is performed by thresholding sigmoid outputs for each label in a multi-label setting (&&&^^^^3query3^^^^&&&).

The training objective is binary cross-entropy with logits. In the notation provided for the system, the loss for the label set is

PRESERVED_PLACEHOLDER_^^^^3query3^^^^

A label PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^ is assigned when its probability exceeds a threshold PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^, and that threshold is itself tuned rather than fixed a priori. Optimization uses AdamW; the training procedure additionally explores label smoothing, weight decay, warm-up steps, learning-rate scheduling, and dynamic batch size (&&&^^^^3query3^^^^&&&).

Hyperparameter optimization is a prominent part of the FITS methodology. The study uses SMAC for Bayesian optimization, with a random forest surrogate and log expected improvement acquisition. The search space includes learning rate from $1 \times 10^{-6}$ to $0.01$ on a log scale, weight decay from ^^^^3query3^^^^.^^^^3query3query3query3all:\3^^^^ to ^^^^3query3^^^^.3, warmup ratio from ^^^^3query3^^^^.^^^^3query3query3query3all:\3^^^^ to ^^^^3query3^^^^.^^^^3all:\3 label smoothing from ^^^^3query3^^^^.^^^^3query3query3query3all:\3^^^^ to ^^^^3query3^^^^.^^^^3all:\3 classification threshold from ^^^^3query3^^^^.3 to ^^^^3query3^^^^.6 in steps of ^^^^3query3^^^^.^^^^3query3all:\3 scheduler type, and training epochs from ^^^^3all:\3^^^^5 to 35. The reported setup uses ^^^^3all:\3query3query3query3^^^^ trials per model and 5 seeds, with model selection driven by validation F^^^^3all:\3^^^^ (&&&^^^^3query3^^^^&&&).

The best reported model is RoBERTa base, which attains weighted F^^^^3all:\3^^^^-score ^^^^3query3^^^^.683 and outperforms all baselines and other transformer variants. Evaluation also includes micro and macro F^^^^3all:\3^^^^. This result is technically notable less because of absolute score magnitude than because it establishes that domain-adapted transformer classifiers can extract issue-level sustainability evidence from a relatively small, curated corpus with better performance than keyword systems or shallow text classifiers (&&&^^^^3query3^^^^&&&).

A plausible implication is that FITS belongs to a broader pattern in applied NLP where domain specificity compensates for limited dataset size. The paper itself emphasizes that scientific-domain and climate-domain pretraining were explicitly tested, indicating that pretraining corpus selection is treated as a first-order modeling variable rather than an incidental implementation detail (&&&^^^^3query3^^^^&&&).

4. Interface design, evidence traceability, and user studies

FITS is implemented as a Flask-based web application. The major functions documented for the interface are search by brand and keyword, interactive exploration of results through a pie chart and table, upload-and-analyze support for user documents, filtering by class, keyword, and brand, and direct source traceability through links to the originating DOI or NGO webpage. The tool also handles upload and processing failures gracefully, and uploaded data is managed with session-based privacy controls (&&&^^^^3query3^^^^&&&).

The interface is important because FITS is not conceived merely as a benchmark classifier. It is a decision-support surface for navigating evidence. Each snippet displayed to the user is linked back to its source, and the system explicitly avoids LLM-style generated claims. In effect, FITS operationalizes a retrieval-plus-classification paradigm in which interpretability comes from evidence provenance rather than post hoc explanation alone (&&&^^^^3query3^^^^&&&).

The tool was evaluated in two focus groups of potential users, with ^^^^3all:\3query3^^^^ participants each. Participants had diverse backgrounds, were regular online shoppers, the sessions were held in German, and the tool itself was in English. Reported positive reactions included strong demand for credible sustainability information, appreciation of source traceability, and approval of the upload functionality. Critiques included the perception that the system was too scientific and not sufficiently consumer friendly, that pie-chart visualizations were confusing, and that navigation, readability, color scheme, search, and translation required improvement. Participants also requested quick comparison features, traffic-light style summaries, saving and export functions, broader integration, and support for analyzing screenshots or URLs (&&&^^^^3query3^^^^&&&).

These findings are methodologically consequential. They indicate that evidence quality alone is insufficient for adoption: presentation, summarization, and comparative framing are also system-critical. This suggests a separation between backend rigor and frontend abstraction. The backend of FITS is deliberately source-conservative; the frontend, however, may need stronger aggregation and simplification layers for non-expert use without compromising evidentiary fidelity.

5. Relation to adjacent fashion-sustainability infrastructures

Although FITS itself is a text-centric system, related research identifies complementary infrastructures that could extend its scope into a more complete sustainability information stack.

At the materials layer, TextileNet provides a taxonomy-based fashion textile dataset with 33 fibres labels, ^^^^3 OR ti:\3^^^^7 fabrics labels, and 76^^^^3query3^^^^,949 images, created in collaboration with material scientists. Baselines with ResNet5^^^^3query3^^^^ and ViT-tiny report top-5 accuracies above 8^^^^3query3^^^^%, and the dataset is explicitly positioned for textile classification, supply-chain optimization, and sustainability-oriented reuse, repair, and recycling workflows (&&&^^^^3all:\3^^^^6&&&). This suggests a possible coupling between textual sustainability evidence in FITS and image-based material recognition pipelines.

At the product-schema layer, GreenDB extends schema.org Product with fields including sustainabilityLabels and labelType, distinguishes third-party from private labels, and is designed for ready integration into existing product catalogs (&&&^^^^3 OR ti:\3^^^^&&&). FITS could therefore be interpreted as a complementary evidence-extraction layer to a retail-product schema layer: GreenDB structures product-side sustainability metadata, while FITS structures brand- and issue-level textual evidence.

At the interoperability layer, DDOIF defines a protocol for Digital Data Exchange and Organisation in Fashion with a dictionary of over ^^^^3all:\3query3query3query3^^^^ fashion classes and subclasses, a binary .ddof file structure, and an API for import/export across the fashion value chain. The protocol is motivated by ambiguity and inconsistency in fashion data representation and is explicitly connected to traceability, transparency, and AI ingestion (&&&^^^^3all:\3^^^^8&&&). In that context, FITS can be seen as operating on sustainability knowledge about brands and supply chains, whereas DDOIF addresses standardized representation of item-level attributes and media.

At the assessment layer, CONFARM provides a four-step procedure—lifecycle documentation, cause-effect mapping, conflict database construction, and multi-criteria scoring—and defines a sustainability ratio

$R = \frac{N}{P+N},$

where lower values indicate higher sustainability. In its fashion-sector demonstration, slow fashion yields lower ratios than fast fashion across conflict densities, while AI-assisted and manual extraction lead to consistent conclusions (Chakrabarti, 12 Dec 2025). FITS does not implement this scoring framework, but related work of this kind suggests one route by which classified evidence could be aggregated into explicit trade-off analyses.

At the design and simulation layer, Textile IR introduces a bidirectional intermediate representation linking manufacturing-valid CAD, physics-based simulation, and lifecycle assessment through a seven-layer Verification Ladder, with explicit uncertainty propagation and real-time sustainability updates under material or pattern changes (&&&^^^^3 OR ti:\3query3^^^^&&&). Together with fit-aware virtual try-on work such as FitGAN and the FIT dataset—which address realistic garment fit, returns reduction, and measurement-grounded visualization—these systems indicate that “fashion information” for sustainability may extend from textual evidence to material identity, interoperability, forecasting, fit, manufacturability, and lifecycle feedback (&&&^^^^3 OR ti:\3all:\3^^^^&&&, &&&^^^^3 OR ti:\3 OR ti:\3^^^^&&&).

6. Limitations, misconceptions, and research directions

The primary limitation of FITS as documented is scope. It classifies sustainability information from credible text; it does not directly verify supply-chain facts, compute lifecycle impacts, infer garment materials from images, or solve end-to-end product assessment. The system’s evidence is constrained by the availability of scientific and NGO documentation, and focus-group participants explicitly encountered data sparsity in the form of “no results for brand X” (&&&^^^^3query3^^^^&&&).

A second limitation is usability stratification. The underlying user study reports that the system was perceived as too scientific by some participants, with requests for clearer summaries and comparison tools. This exposes a recurring tension in sustainability informatics: retaining issue-level nuance while avoiding interfaces that overwhelm non-specialist users. FITS currently prioritizes traceability and evidence fidelity over aggressive summarization (&&&^^^^3query3^^^^&&&).

A third limitation is that the corpus is intentionally not frozen as a static public dataset because fast-changing brand data raises misinformation risks. This is methodologically defensible, but it means reproducibility depends on reconstruction procedures and source-list maintenance rather than a fixed benchmark snapshot (&&&^^^^3query3^^^^&&&).

Several future directions emerge from adjacent literature. Predictive analytics for New Fashion Products Performance Forecasting has been surveyed as a route to reducing overproduction, waste, and inefficient resource use, with multimodal approaches integrating tabular, visual, textual, and exogenous signals such as Google Trends (&&&^^^^3 OR ti:\3^^^^6&&&). Computer-vision surveys also identify material recognition, parsing, retrieval, attribute recognition, and recommendation as established technical components that could be coupled with sustainability-oriented interfaces (&&&^^^^3 OR ti:\3^^^^7&&&). A plausible implication is that a mature FITS ecosystem would not remain purely textual: it would connect evidence extraction with standardized product schemas, material taxonomies, lifecycle-aware CAD, and demand forecasting.

In that expanded sense, FITS denotes both a specific ^^^^3 OR ti:\3query3 OR ti:\3^^^^5 prototype and a broader systems agenda. The prototype establishes a domain-adapted, source-traceable NLP core for fashion sustainability information. Related work shows how such a core could be integrated with product databases, interoperable data protocols, textile-recognition datasets, conflict-mapping frameworks, and physics-aware design representations. The central research problem remains unchanged across these layers: sustainability information in fashion must be machine-readable, evidence-linked, and sufficiently structured to support decision-making without collapsing complex trade-offs into opaque or hallucinated outputs (&&&^^^^3query3^^^^&&&).