DATASHEET.md

# Datasheet for PCFBench

This datasheet follows the *Datasheets for Datasets* template
(Gebru et al., 2021). It documents PCFBench, a process-based product
carbon footprint benchmark for evaluating LLMs and agents on the
operational steps of life-cycle assessment (LCA).

## 1. Motivation

**For what purpose was the dataset created?**

To evaluate whether language models and tool-using agents can perform
the operational steps of a process-based product carbon footprint
(PCF) calculation: decomposing a product into its bill of materials,
deciding when to map vs. further decompose, mapping foreground
materials onto an ecoinvent process, extracting physical input rates
from technical documentation, and producing a single-shot total
kgCO₂e prediction that can be checked against an EPD ground truth.
Existing sustainability benchmarks are predominantly spend-based
(emissions ≈ \$ × industry-average factor); PCFBench targets the
process-based operational pipeline, where a wrong choice at any step
silently propagates into the final number.

**Who created the dataset and on behalf of which entity?**

Krishna Rao, Andrew Dumit, Shaena Ulissi, Jacob Feintzeig, P. James
Joyce, Daniel Frank, Steven Watson, Jonathan Glidden, Gizem Ilayda Dinc
and Travis M. Kwee, on behalf of Watershed Technology, Inc.

**Who funded the creation of the dataset?**

Watershed Technology, Inc.

## 2. Composition

**What do the instances that comprise the dataset represent?**

Six per-task JSONL files plus one shared candidate-set JSONL. Each
task instance is a single evaluation item with `input` (what the
model receives), `expected_output` (the ground-truth target), and
`metadata` (stratification + provenance fields).

| File | Items | What an instance is |
| --- | ---: | --- |
| `task1_decomposition.jsonl`       |    94 | A product (name + description + unit) annotated with the expert-decided list of bill-of-materials components. |
| `task2_triage.jsonl`              |   200 | A market node + a material's parent context (description, market name) annotated with the binary decision *should_map* (resolve to a single ecoinvent process) vs. decompose further. |
| `task3_mapping.jsonl`             |   109 | A foreground material (name + optional description / supplier / purchaser context) annotated with one or more acceptable ecoinvent reference products, plus optional relevant- and banned-substring rules. |
| `task4_extraction_material.jsonl` |    22 | A free-form question over a real technical document, annotated with the set of material-input claims (value, unit, verbatim evidence quotes) that the document supports. |
| `task5_extraction_energy.jsonl`   |    14 | Same as Task 4, but for energy-input claims. |
| `task7_epd.jsonl`                 |   175 | An EPD-described product (name + description + optional composition + region + recycled content) annotated with the EPD's published cradle-to-gate kgCO₂e value. |

**How many instances are there in total?**

614 task items across 6 evaluation tasks. Tasks 4 and 5 are scored at
the **claim** level — there are 89 ground-truth claims across the 36
extraction documents (55 material + 34 energy).

**Does the dataset contain all possible instances or is it a sample?**

Each task is a curated benchmark slice, not an exhaustive enumeration:

- **Task 1**: the 94 EPD-derived products with explicit composition,
  out of an initial 187 EPD-sourced items (composition-bearing subset).
- **Task 2**: 200 items hand-annotated by sustainability-domain
  experts (100 *should_map* + 100 *should_decompose*), stratified
  across product categories and capped at 8 items per root material
  per bucket.
- **Task 3**: 109 items hand-curated by sustainability experts from a
  larger 1,198-item pool, retaining single-activity mappings only
  (composite forming-process + raw-material pairs are filtered out).
- **Tasks 4–5**: 36 documents drawn from peer-reviewed LCA
  literature, EPD source documents, and industry technical
  specifications, after a per-annotator-latest +
  cross-annotator-unanimous-on-include voting rule eliminated
  duplicative and irrelevant claims. CO₂-intensity items were pruned
  upstream; 3 documents archived because every claim had at least one
  dissenting annotator.
- **Task 7**: 175 EPDs from environdec.com spanning 11+ environdec
  product categories. Selection biased toward kgCO₂e in the
  0.06 – 19,600 range to span the 5-orders-of-magnitude footprint
  variability practitioners encounter.

**Is each instance labeled?**

Yes. Every instance has `expected_output` populated. Tasks 4–5
additionally carry verbatim evidence quotes per claim, validated as
substrings of the source `document_text` after whitespace
normalization (0/227 substring failures at last audit).

**Are there recommended data splits?**

The full dataset is treated as a held-out evaluation set; there is no
designated train split. Reviewers are expected to evaluate without
training-on-test contamination. We commit to versioned releases
(starting `v1.0`) and recommend pinning to a release tag when
reporting numbers.

**Are there known errors, sources of noise, or redundancies in the dataset?**

- Tasks 4–5 ground truth is the unanimous-on-include consensus of up
  to 3 LCA practitioners. Single-annotator edge cases — typically
  numbers that are technically present in the source but not the
  document's primary measured quantity — are dropped, which biases
  the GT toward the document's most prominent reported claims.
- Task 1 scoring relies on an LLM judge (Gemini 2.5 Flash) that
  produces compositional match groups; precision/recall/F₁ are
  computed deterministically from those groups. A 50-item human
  audit on the Gemini 3.1 Pro run found 88% agreement (Wilson 95% CI
  [0.76, 0.94]) with one author's hand grades; the bias on the
  headline F₁ is approximately +0.01 absolute (paper F₁ 0.735 →
  human-aligned ~0.745) and does not reorder the top-3 models.
- The Tasks 2/3 candidate set is region-agnostic by construction —
  geographic variability is a known limitation flagged for future work.

**Is the dataset self-contained, or does it link to external resources?**

Self-contained for scoring. Each task JSONL embeds everything needed
to score predictions: ground-truth values, unit info, evidence
quotes, substrings, and category labels. Tasks 4–5 include the
OCR-extracted `document_text` directly in each row; Tasks 4–5 source
documents and the Task 7 EPDs additionally carry `source_url` for
provenance. Tasks 2 and 3 require an ecoinvent v3.11 candidate set
(2,574 rows) at evaluation time; that candidate set is not bundled
in this dataset and ships instead with the companion code repository.

**Does the dataset contain confidential, sensitive, or PII?**

No. All source materials are publicly disclosed Environmental
Product Declarations, peer-reviewed LCA literature, and the
public ecoinvent v3.11 Database Overview workbook.
No personally identifiable information is present.

## 3. Collection Process

**How was the data associated with each instance acquired?**

- **Task 1 / Task 7**: parsed from publicly disclosed EPDs at
  environdec.com; `product_name`, `description`, `composition`,
  `geography`, `recycled_content`, and `kgco2e` are extracted from
  the EPD's structured PDF fields. Inclusion of these parsed
  metadata fields in PCFBench is covered by a written permission
  grant from EPD International. Source EPDs remain
  accessible at environdec.com via each row's `source_url`. See the
  accompanying paper's Appendix O for the full permission letter.
- **Task 2**: 200 hand-labelled map-vs-decompose decision points,
  balanced 100/100. Each item carries a sustainability-expert
  decision of *should_map* vs. *should_decompose*.
- **Task 3**: 109 material-to-ecoinvent mappings hand-curated by
  domain experts, each with a primary reference product, ranked
  alternatives, and a banned-substring of known-wrong matches.
- **Tasks 4–5**: each annotator independently submitted candidate
  claims with verbatim source quotes; a per-annotator latest-decision
  pass dedupes within annotator, then the cross-annotator
  unanimous-on-include rule keeps a claim only if no annotator
  marked it irrelevant or excluded; (rounded value, normalized unit)
  collapses claims within a document and unions their evidence
  quotes; every retained quote is asserted to be a verbatim substring
  of the source `document_text` after whitespace normalization.

**What mechanisms or procedures were used to collect the data?**

For Tasks 1 and 7: scripted extraction from public sources
(environdec EPD PDFs). For Tasks 2 and 3: structured human annotation by
sustainability-domain experts working from real material / market
contexts. For Tasks 4–5: a Streamlit annotation interface where
domain experts review LLM-proposed candidate claims, with mandatory
verbatim-evidence selection from the source document.

**Who was involved in the data collection process?**

Sustainability-domain experts and LCA practitioners employed by the
authors' organization performed annotation and curation, plus the
authors performed dataset assembly, judge prompt engineering, and
audit. Six annotators participated, all from the LCA and
sustainability domain; some hold PhDs in that field.

**Over what timeframe was the data collected?**

The benchmark slices were assembled over February 2025–April 2026.
Underlying source materials (EPDs, literature, ecoinvent v3.11)
predate this window.

## 4. Preprocessing / Cleaning / Labeling

**Was any preprocessing/cleaning/labeling of the data done?**

Yes:

- **Composite-mapping filter** (Task 3): items whose ground truth
  required two ecoinvent activities (e.g., a forming process plus a
  raw material) are excluded; 0/109 retained items have
  `reference_product_2` non-null after the filter.
- **Unanimous-on-include consensus** (Tasks 4–5): each annotator's
  latest decision is taken, then claims are kept only if every
  annotator with an opinion voted *include* (or *duplicative*, which
  we treat as include).
- **Within-document (value, unit) dedup** (Tasks 4–5): claims that
  share a (rounded value, normalized unit) are collapsed and their
  evidence lists unioned. This intentionally destroys the
  per-`parameter_name` distinction — a model that produces the right
  number with the right unit is correct, regardless of which named
  parameter the document called it.
- **Item-level material/energy classification** (Tasks 4–5): each
  item is tagged as *material* or *energy* based on the question
  text, partitioning the 36-item dataset into the two task files
  (22 + 14).
- **Product-category tagging** (all tasks): each item is assigned
  one of 12 environdec product categories for stratified analysis.

**Was the "raw" data saved in addition to the preprocessed/cleaned/labeled data?**

The published JSONL files are the post-cleaning canonical release.
Intermediate annotator-decision records (per-claim include / exclude /
duplicative votes for Tasks 4–5) are retained by the authors and can
be made available on request for replication of the consensus
construction.  The build scripts that produce the JSONLs are
open-sourced in the companion code repository.

## 5. Uses

**Has the dataset been used for any tasks already?**

No. PCFBench is being released for the first time alongside this
submission; the accompanying paper is its first published use.

**What other tasks could the dataset be used for?**

LCA-specific:
- Fine-tuning evaluation for sustainability-focused agents
- RL training environment with verifiable per-step reward shaping
- Studying compositional vs. monolithic LCA pipelines

Generic LLM / agent capability probes:
- Top-down decomposition vs. single-shot total estimation: Tasks 1 + 7
  share the same underlying products, so models can be probed on
  whether they reason consistently across BOM expansion (top-down)
  and total kgCO₂e prediction (single-shot, no intermediate decomposition
  exposed) on the same physical components.
- Compositional error attribution: Tasks 1, 3, 4, 5, and 7 form a
  compositional pipeline, allowing residual error to be attributed
  to specific stages instead of just observed at the aggregate-output
  level.
- Evidence-grounded extraction with verbatim quote constraints,
  beyond the LCA domain.
- Numerical reasoning over multi-unit physical quantities
  (mass-fraction, energy intensity, power, temperature, time)
  with required unit-correctness.
- Calibrated abstention vs. fabrication under document-vs-prior
  uncertainty (the Tasks 4–5 query-only ablation is one such
  setup).
- Semantic matching across heterogeneous ontologies (foreground
  material name → ecoinvent reference product) under acronym /
  abbreviation / foreign-language / vague-input shifts.
- Tool-use vs. in-context retrieval ablation: each of Tasks 2 and
  3 has both a single-shot and an agentic (tool-using) variant
  on the same underlying items.
- Auditing LLM calibration on order-of-magnitude reasoning under
  underspecification.

**Is there anything about the composition of the dataset or the way it was collected and preprocessed/cleaned/labeled that might impact future uses?**

- The Tasks 2/3 candidate set is region-agnostic; geographic
  stratification of Tasks 2 + 3 is a known future-work item.
- Tasks 4 + 5 sources skew toward industrial-process literature
  (extruders, furnaces, polymer processing); generalization to other
  process classes is untested at this scale.
- The Task 1 judge is itself an LLM (Gemini 2.5 Flash). We've
  validated agreement with a human author at 88% on a 50-item
  stratified sample; treat this as a strong but imperfect signal.

**Are there tasks for which the dataset should not be used?**

- **Direct use in regulatory PCF reporting.** This is an evaluation
  benchmark; numbers produced against it should not be substituted
  for an audited LCA.
- **Greenwashing.** Models that pass PCFBench's metrics should not be
  marketed as "verified LCA tools." Achieving high accuracy on the
  benchmark is necessary but not sufficient for production
  deployment.

## 6. Distribution

**Will the dataset be distributed to third parties outside of the entity on behalf of which the dataset was created?**

Yes. The dataset is publicly distributed via Hugging Face and
licensed under CC BY-NC-SA 4.0 (see `LICENSE`).

**How will the dataset be distributed?**

- Hugging Face dataset repository:
  `https://huggingface.co/datasets/Watershed-Climate/PCFBench`
- Croissant metadata is generated by Hugging Face from the dataset card
  and served at the repository's `/croissant` endpoint.
- Companion code repository:
  `https://github.com/watershed-climate/pcfbench`

**When will the dataset be distributed?**

Publicly available.

**Will the dataset be distributed under a copyright or other intellectual property (IP) license?**

CC BY-NC-SA 4.0 for the dataset as a whole.

## 7. Maintenance

**Who will be supporting/hosting/maintaining the dataset?**

The authors, with versioned releases on Hugging Face.

**Is there an erratum?**

Will be tracked in the dataset's Community tab on Hugging Face and
via `CHANGELOG.md` in the companion code repository.

**Will the dataset be updated?**

Yes — point releases will be issued when:

- Annotation errors are reported and corrected
- The companion code repository's ecoinvent candidate set is
  regenerated for a new ecoinvent release
- Additional product-category coverage (notably *Infrastructure &
  buildings*) is added

When reporting numbers we recommend pinning to a release tag.

**If others want to extend/augment/build on/contribute to the dataset, is there a mechanism for them to do so?**

Yes — open a discussion in the dataset's Community tab on Hugging
Face, or a pull request against the companion code repository.