README.md

---
pretty_name: PCFBench
license: cc-by-nc-sa-4.0
language:
  - en
size_categories:
  - n<1K
task_categories:
  - question-answering
  - text-classification
tags:
  - arxiv:2608.27716
  - life-cycle-assessment
  - carbon-footprint
  - sustainability
  - agents
  - benchmark
configs:
  - config_name: task1_decomposition
    data_files: task1_decomposition.jsonl
  - config_name: task2_triage
    data_files: task2_triage.jsonl
  - config_name: task3_mapping
    data_files: task3_mapping.jsonl
  - config_name: task4_extraction_material
    data_files: task4_extraction_material.jsonl
  - config_name: task5_extraction_energy
    data_files: task5_extraction_energy.jsonl
  - config_name: task7_epd
    data_files: task7_epd.jsonl
---

# PCFBench

Paper: [arXiv:2608.27716](https://arxiv.org/abs/2608.27716) · Code: [watershed-climate/pcfbench](https://github.com/watershed-climate/pcfbench)

Process-based Product Carbon Footprint benchmark for evaluating LLMs and
agents on the operational steps of life-cycle assessment (LCA): bill-of-
materials decomposition, mapping triage, ecoinvent process matching,
literature extraction of physical input rates, and total kgCO₂e
prediction against expert-grounded EPDs.

## Tasks

| ID | Task                                | Items | GT claims | Headline metric |
| -- | ----------------------------------- | ----: | --------: | --------------- |
| 1  | Product decomposition (BOM)         |    94 |        94 | Judge-aligned F₁ on compositional match groups |
| 2  | Mapping triage                      |   200 |       200 | Accuracy / F₁ on `should_map` binary |
| 3  | Background-database mapping         |   109 |       109 | Exact-match top-1 against expert reference products |
| 4  | Material input-rate extraction      |    22 |        55 | Claim F₁ on greedy (value, unit) match |
| 5  | Energy input-rate extraction        |    14 |        34 | Claim F₁ on greedy (value, unit) match |
| 7  | Total kgCO₂e prediction (EPD)       |   175 |       175 | Median \|RE\|, within-2× / within-5× rate |

(Step 6 is deterministic arithmetic and not separately evaluated.)

## Files

| File | Rows | Description |
| --- | ---: | --- |
| `task1_decomposition.jsonl`         |    94 | Step 1 — product → BOM. `expected_output.components: list[str]`. |
| `task2_triage.jsonl`                |   200 | Step 2 — given a market node + material context, decide map vs. decompose. `expected_output.should_map: bool`. |
| `task3_mapping.jsonl`               |   109 | Step 3 — material → ecoinvent reference product. `expected_output.options: list[str]` (composite-mapping items already filtered out). |
| `task4_extraction_material.jsonl`   |    22 | Step 4 — extract material input rates from a technical document. `expected_output.claims: list[{value, unit, evidence}]`. |
| `task5_extraction_energy.jsonl`     |    14 | Step 5 — extract energy input rates. Same shape as Task 4. |
| `task7_epd.jsonl`                   |   175 | Step 7 — single-shot total kgCO₂e prediction against an EPD ground truth. `expected_output.kgco2e: float`. |

## Schema

Every task row uses the same envelope:

```json
{
  "id": "...",
  "input": { /* task-specific */ },
  "expected_output": { /* task-specific ground truth */ },
  "metadata": {
    "product_category": "Metal, mineral, plastic & glass products",
    /* task-specific extras: vagueness_severity, request_id, tags=[n_components_*], ... */
  }
}
```

The 12 environdec-aligned product categories used across tasks are:
*Chemical products, Construction products, Electricity / steam / fuels,
Food & beverages, Furniture & other goods, Infrastructure & buildings,
Machinery & equipment, Metal, mineral, plastic & glass products, Paper
and plastic products, Services, Textiles, footwear & apparel, Vehicles
& transport equipment.* All tasks except Tasks 4–5 currently miss
**Infrastructure & buildings**; Tasks 4–5 cover a subset because the
extraction documents are concentrated in industrial-process literature.

See `DATASHEET.md` for the full Datasheet for Datasets, including
collection process, annotation protocol, intended uses, and limitations.

## Loading

Each task is a separate config:

```python
from datasets import load_dataset

decomposition = load_dataset("Watershed-Climate/PCFBench", "task1_decomposition", split="train")
epd = load_dataset("Watershed-Climate/PCFBench", "task7_epd", split="train")
```

The files are plain JSONL, so `pandas.read_json(..., lines=True)` or
`json.loads` per line work equally well on a local copy.

The companion code repository at
[watershed-climate/pcfbench](https://github.com/watershed-climate/pcfbench)
provides a ready-to-run eval harness for all 6 task variants.

## Citation

```bibtex
@misc{pcfbench2026,
  title  = {PCFBench: A Diagnostic Benchmark for Product Carbon Footprint Estimation},
  author = {Rao, Krishna and Dumit, Andrew and Ulissi, Shaena and
            Feintzeig, Jacob and Joyce, P. James and Frank, Daniel and
            Watson, Steven and Glidden, Jonathan and Dinc, Gizem Ilayda and
            Kwee, Travis M.},
  year   = {2026},
  eprint = {2608.27716},
  archivePrefix = {arXiv}
}
```

## License

[CC BY-NC-SA 4.0](LICENSE).

`task7_epd.jsonl` references publicly disclosed Environmental Product
Declarations from environdec.com; each item retains the original
EPD's `source_url`.