---
license: mit
language:
  - en
  - de
  - fr
  - it
  - es
  - pt
  - hy
  - ka
  - ne
  - sr
  - zh
  - fil
  - nl
pretty_name: "LEGEX Inference Results: 4 Systems, 6 Runs on 19 Jurisdictions"
size_categories:
  - 1K<n<10K
task_categories:
  - text-classification
  - token-classification
tags:
  - legal
  - benchmark
  - llm-evaluation
  - information-extraction
  - inference-outputs
dataset_info:
  - config_name: harvey
    features:
      - name: case_id
        dtype: string
      - name: link
        dtype: string
      - name: legal_subject_judgement
        dtype: string
      - name: trial_start_date
        dtype: date32
      - name: trial_end_date
        dtype: date32
      - name: dispute_value_nominal
        dtype: string
      - name: Currency_dispute_value_nominal
        dtype: string
      - name: plaintiff_loosing_share
        dtype: string
      - name: court_cost_awarded_nominal
        dtype: string
      - name: Currency_court_cost_awarded_nominal
        dtype: string
      - name: party_compensation_awarded_nominal
        dtype: string
      - name: Currency_party_compensation_awarded_nominal
        dtype: string
      - name: plaintiffs_all_count
        dtype: string
      - name: defendants_all_count
        dtype: string
      - name: plaintiff_no1_ISIC1_industry_category
        dtype: string
      - name: defendant_no1_ISIC1_industry_category
        dtype: string
      - name: model
        dtype: string
      - name: inference_date
        dtype: string
      - name: error
        dtype: string
      - name: comment
        dtype: string
      - name: original_input
        dtype: string
  - config_name: harvey_2
    features:
      - name: case_id
        dtype: string
      - name: link
        dtype: string
      - name: legal_subject_judgement
        dtype: string
      - name: trial_start_date
        dtype: date32
      - name: trial_end_date
        dtype: date32
      - name: dispute_value_nominal
        dtype: string
      - name: Currency_dispute_value_nominal
        dtype: string
      - name: plaintiff_loosing_share
        dtype: string
      - name: court_cost_awarded_nominal
        dtype: string
      - name: Currency_court_cost_awarded_nominal
        dtype: string
      - name: party_compensation_awarded_nominal
        dtype: string
      - name: Currency_party_compensation_awarded_nominal
        dtype: string
      - name: plaintiffs_all_count
        dtype: string
      - name: defendants_all_count
        dtype: string
      - name: plaintiff_no1_ISIC1_industry_category
        dtype: string
      - name: defendant_no1_ISIC1_industry_category
        dtype: string
      - name: model
        dtype: string
      - name: inference_date
        dtype: string
      - name: error
        dtype: string
      - name: comment
        dtype: string
      - name: original_input
        dtype: string
  - config_name: gemini
    features:
      - name: case_id
        dtype: string
      - name: link
        dtype: string
      - name: legal_subject_judgement
        dtype: string
      - name: trial_start_date
        dtype: date32
      - name: trial_end_date
        dtype: date32
      - name: dispute_value_nominal
        dtype: string
      - name: Currency_dispute_value_nominal
        dtype: string
      - name: plaintiff_loosing_share
        dtype: string
      - name: court_cost_awarded_nominal
        dtype: string
      - name: Currency_court_cost_awarded_nominal
        dtype: string
      - name: party_compensation_awarded_nominal
        dtype: string
      - name: Currency_party_compensation_awarded_nominal
        dtype: string
      - name: plaintiffs_all_count
        dtype: string
      - name: defendants_all_count
        dtype: string
      - name: plaintiff_no1_ISIC1_industry_category
        dtype: string
      - name: defendant_no1_ISIC1_industry_category
        dtype: string
      - name: model
        dtype: string
      - name: inference_date
        dtype: string
      - name: error
        dtype: string
      - name: comment
        dtype: string
      - name: original_input
        dtype: string
  - config_name: gpt
    features:
      - name: case_id
        dtype: string
      - name: link
        dtype: string
      - name: legal_subject_judgement
        dtype: string
      - name: trial_start_date
        dtype: date32
      - name: trial_end_date
        dtype: date32
      - name: dispute_value_nominal
        dtype: string
      - name: Currency_dispute_value_nominal
        dtype: string
      - name: plaintiff_loosing_share
        dtype: string
      - name: court_cost_awarded_nominal
        dtype: string
      - name: Currency_court_cost_awarded_nominal
        dtype: string
      - name: party_compensation_awarded_nominal
        dtype: string
      - name: Currency_party_compensation_awarded_nominal
        dtype: string
      - name: plaintiffs_all_count
        dtype: string
      - name: defendants_all_count
        dtype: string
      - name: plaintiff_no1_ISIC1_industry_category
        dtype: string
      - name: defendant_no1_ISIC1_industry_category
        dtype: string
      - name: model
        dtype: string
      - name: inference_date
        dtype: string
      - name: error
        dtype: string
      - name: comment
        dtype: string
      - name: original_input
        dtype: string
  - config_name: legora_1
    features:
      - name: case_id
        dtype: string
      - name: link
        dtype: string
      - name: legal_subject_judgement
        dtype: string
      - name: trial_start_date
        dtype: date32
      - name: trial_end_date
        dtype: date32
      - name: dispute_value_nominal
        dtype: string
      - name: Currency_dispute_value_nominal
        dtype: string
      - name: plaintiff_loosing_share
        dtype: string
      - name: court_cost_awarded_nominal
        dtype: string
      - name: Currency_court_cost_awarded_nominal
        dtype: string
      - name: party_compensation_awarded_nominal
        dtype: string
      - name: Currency_party_compensation_awarded_nominal
        dtype: string
      - name: plaintiffs_all_count
        dtype: string
      - name: defendants_all_count
        dtype: string
      - name: plaintiff_no1_ISIC1_industry_category
        dtype: string
      - name: defendant_no1_ISIC1_industry_category
        dtype: string
      - name: model
        dtype: string
      - name: inference_date
        dtype: string
      - name: error
        dtype: string
      - name: comment
        dtype: string
      - name: original_input
        dtype: string
  - config_name: legora_2
    features:
      - name: case_id
        dtype: string
      - name: link
        dtype: string
      - name: legal_subject_judgement
        dtype: string
      - name: trial_start_date
        dtype: date32
      - name: trial_end_date
        dtype: date32
      - name: dispute_value_nominal
        dtype: string
      - name: Currency_dispute_value_nominal
        dtype: string
      - name: plaintiff_loosing_share
        dtype: string
      - name: court_cost_awarded_nominal
        dtype: string
      - name: Currency_court_cost_awarded_nominal
        dtype: string
      - name: party_compensation_awarded_nominal
        dtype: string
      - name: Currency_party_compensation_awarded_nominal
        dtype: string
      - name: plaintiffs_all_count
        dtype: string
      - name: defendants_all_count
        dtype: string
      - name: plaintiff_no1_ISIC1_industry_category
        dtype: string
      - name: defendant_no1_ISIC1_industry_category
        dtype: string
      - name: model
        dtype: string
      - name: inference_date
        dtype: string
      - name: error
        dtype: string
      - name: comment
        dtype: string
      - name: original_input
        dtype: string
configs:
  - config_name: harvey
    data_files:
      - split: armenia
        path: data/am/inference_harvey.jsonl
      - split: australia
        path: data/au/inference_harvey.jsonl
      - split: belgium
        path: data/be/inference_harvey.jsonl
      - split: brazil
        path: data/br/inference_harvey.jsonl
      - split: france
        path: data/fr/inference_harvey.jsonl
      - split: georgia
        path: data/ge/inference_harvey.jsonl
      - split: germany
        path: data/de/inference_harvey.jsonl
      - split: hong_kong
        path: data/hk/inference_harvey.jsonl
      - split: india
        path: data/in/inference_harvey.jsonl
      - split: nepal
        path: data/np/inference_harvey.jsonl
      - split: new_zealand
        path: data/nz/inference_harvey.jsonl
      - split: philippines
        path: data/ph/inference_harvey.jsonl
      - split: serbia
        path: data/rs/inference_harvey.jsonl
      - split: singapore
        path: data/sg/inference_harvey.jsonl
      - split: spain
        path: data/es/inference_harvey.jsonl
      - split: switzerland
        path: data/ch/inference_harvey.jsonl
      - split: taiwan
        path: data/tw/inference_harvey.jsonl
      - split: united_kingdom
        path: data/uk/inference_harvey.jsonl
      - split: united_states
        path: data/us/inference_harvey.jsonl
  - config_name: harvey_2
    data_files:
      - split: armenia
        path: data/am/inference_harvey_2.jsonl
      - split: australia
        path: data/au/inference_harvey_2.jsonl
      - split: belgium
        path: data/be/inference_harvey_2.jsonl
      - split: brazil
        path: data/br/inference_harvey_2.jsonl
      - split: france
        path: data/fr/inference_harvey_2.jsonl
      - split: georgia
        path: data/ge/inference_harvey_2.jsonl
      - split: germany
        path: data/de/inference_harvey_2.jsonl
      - split: hong_kong
        path: data/hk/inference_harvey_2.jsonl
      - split: india
        path: data/in/inference_harvey_2.jsonl
      - split: nepal
        path: data/np/inference_harvey_2.jsonl
      - split: new_zealand
        path: data/nz/inference_harvey_2.jsonl
      - split: philippines
        path: data/ph/inference_harvey_2.jsonl
      - split: serbia
        path: data/rs/inference_harvey_2.jsonl
      - split: singapore
        path: data/sg/inference_harvey_2.jsonl
      - split: spain
        path: data/es/inference_harvey_2.jsonl
      - split: switzerland
        path: data/ch/inference_harvey_2.jsonl
      - split: taiwan
        path: data/tw/inference_harvey_2.jsonl
      - split: united_kingdom
        path: data/uk/inference_harvey_2.jsonl
      - split: united_states
        path: data/us/inference_harvey_2.jsonl
  - config_name: gemini
    data_files:
      - split: armenia
        path: data/am/inference_gemini.jsonl
      - split: australia
        path: data/au/inference_gemini.jsonl
      - split: belgium
        path: data/be/inference_gemini.jsonl
      - split: brazil
        path: data/br/inference_gemini.jsonl
      - split: france
        path: data/fr/inference_gemini.jsonl
      - split: georgia
        path: data/ge/inference_gemini.jsonl
      - split: germany
        path: data/de/inference_gemini.jsonl
      - split: hong_kong
        path: data/hk/inference_gemini.jsonl
      - split: india
        path: data/in/inference_gemini.jsonl
      - split: nepal
        path: data/np/inference_gemini.jsonl
      - split: new_zealand
        path: data/nz/inference_gemini.jsonl
      - split: philippines
        path: data/ph/inference_gemini.jsonl
      - split: serbia
        path: data/rs/inference_gemini.jsonl
      - split: singapore
        path: data/sg/inference_gemini.jsonl
      - split: spain
        path: data/es/inference_gemini.jsonl
      - split: switzerland
        path: data/ch/inference_gemini.jsonl
      - split: taiwan
        path: data/tw/inference_gemini.jsonl
      - split: united_kingdom
        path: data/uk/inference_gemini.jsonl
      - split: united_states
        path: data/us/inference_gemini.jsonl
  - config_name: gpt
    data_files:
      - split: armenia
        path: data/am/inference_gpt.jsonl
      - split: australia
        path: data/au/inference_gpt.jsonl
      - split: belgium
        path: data/be/inference_gpt.jsonl
      - split: brazil
        path: data/br/inference_gpt.jsonl
      - split: france
        path: data/fr/inference_gpt.jsonl
      - split: georgia
        path: data/ge/inference_gpt.jsonl
      - split: germany
        path: data/de/inference_gpt.jsonl
      - split: hong_kong
        path: data/hk/inference_gpt.jsonl
      - split: india
        path: data/in/inference_gpt.jsonl
      - split: nepal
        path: data/np/inference_gpt.jsonl
      - split: new_zealand
        path: data/nz/inference_gpt.jsonl
      - split: philippines
        path: data/ph/inference_gpt.jsonl
      - split: serbia
        path: data/rs/inference_gpt.jsonl
      - split: singapore
        path: data/sg/inference_gpt.jsonl
      - split: spain
        path: data/es/inference_gpt.jsonl
      - split: switzerland
        path: data/ch/inference_gpt.jsonl
      - split: taiwan
        path: data/tw/inference_gpt.jsonl
      - split: united_kingdom
        path: data/uk/inference_gpt.jsonl
      - split: united_states
        path: data/us/inference_gpt.jsonl
  - config_name: legora_1
    data_files:
      - split: armenia
        path: data/am/inference_legora_1.jsonl
      - split: australia
        path: data/au/inference_legora_1.jsonl
      - split: belgium
        path: data/be/inference_legora_1.jsonl
      - split: brazil
        path: data/br/inference_legora_1.jsonl
      - split: france
        path: data/fr/inference_legora_1.jsonl
      - split: georgia
        path: data/ge/inference_legora_1.jsonl
      - split: germany
        path: data/de/inference_legora_1.jsonl
      - split: hong_kong
        path: data/hk/inference_legora_1.jsonl
      - split: india
        path: data/in/inference_legora_1.jsonl
      - split: nepal
        path: data/np/inference_legora_1.jsonl
      - split: new_zealand
        path: data/nz/inference_legora_1.jsonl
      - split: philippines
        path: data/ph/inference_legora_1.jsonl
      - split: serbia
        path: data/rs/inference_legora_1.jsonl
      - split: singapore
        path: data/sg/inference_legora_1.jsonl
      - split: spain
        path: data/es/inference_legora_1.jsonl
      - split: switzerland
        path: data/ch/inference_legora_1.jsonl
      - split: taiwan
        path: data/tw/inference_legora_1.jsonl
      - split: united_kingdom
        path: data/uk/inference_legora_1.jsonl
      - split: united_states
        path: data/us/inference_legora_1.jsonl
  - config_name: legora_2
    data_files:
      - split: armenia
        path: data/am/inference_legora_2.jsonl
      - split: australia
        path: data/au/inference_legora_2.jsonl
      - split: belgium
        path: data/be/inference_legora_2.jsonl
      - split: brazil
        path: data/br/inference_legora_2.jsonl
      - split: france
        path: data/fr/inference_legora_2.jsonl
      - split: georgia
        path: data/ge/inference_legora_2.jsonl
      - split: germany
        path: data/de/inference_legora_2.jsonl
      - split: hong_kong
        path: data/hk/inference_legora_2.jsonl
      - split: india
        path: data/in/inference_legora_2.jsonl
      - split: nepal
        path: data/np/inference_legora_2.jsonl
      - split: new_zealand
        path: data/nz/inference_legora_2.jsonl
      - split: philippines
        path: data/ph/inference_legora_2.jsonl
      - split: serbia
        path: data/rs/inference_legora_2.jsonl
      - split: singapore
        path: data/sg/inference_legora_2.jsonl
      - split: spain
        path: data/es/inference_legora_2.jsonl
      - split: switzerland
        path: data/ch/inference_legora_2.jsonl
      - split: taiwan
        path: data/tw/inference_legora_2.jsonl
      - split: united_kingdom
        path: data/uk/inference_legora_2.jsonl
      - split: united_states
        path: data/us/inference_legora_2.jsonl
---

# LEGEX Inference Results

System outputs for the LEGEX benchmark, six review-table extraction
runs (four systems; Harvey and Legora twice each) on the case packets of
all 19 jurisdictions:

`am, au, be, br, ch, de, es, fr, ge, hk, in, np, nz, ph, rs, sg, tw, uk, us`

## The systems

| Abbreviation | `model` field value | Notes |
|--------------|---|---|
| `harvey`     | `harvey` | Harvey Vault Review, a commercial review-table product. Outputs exported from the production system on **2026-05-18** and **2026-06-30** (each record's `inference_date` says which export it came from) — the run reported in the paper. The per-column prompts as configured for each export are in [`prompts/prompts_harvey.jsonl`](prompts/prompts_harvey.jsonl). Harvey is a deployed product that changes over time, so these results are a dated snapshot. |
| `harvey_2`   | `harvey-2` | A second, independent Harvey Vault Review run, exported **2026-08-05**, with the prompts re-entered from the same rulebook (and rephrased by the platform on entry — see [`prompts/prompts_harvey_2.jsonl`](prompts/prompts_harvey_2.jsonl)). Released for robustness comparisons across time and prompt variants; not part of the paper's reported numbers. |
| `gemini`     | `gemini/gemini-3.1-flash-lite` | Schema-constrained pipeline via the Gemini API. |
| `gpt`        | `gpt-5.4-mini` | Schema-constrained pipeline via the OpenAI API. |
| `legora_1`   | `legora-1` | Legora Tabular Review, a commercial review-table product. Exported from the production system on **2026-08-01**; the export contained the full question set twice, i.e. two independent runs — this is the first (left) column group, the run reported in the paper. The per-column prompts as run are in [`prompts/prompts_legora_1.jsonl`](prompts/prompts_legora_1.jsonl). Like Harvey, a dated snapshot of a deployed product. |
| `legora_2`   | `legora-2` | The second (right) column group of the same 2026-08-01 Legora export — an independent run with independently re-entered per-column prompts, see [`prompts/prompts_legora_2.jsonl`](prompts/prompts_legora_2.jsonl). Released for robustness comparisons across prompt variants. |

The `gemini` and `gpt` pipelines received the identical single system prompt
(`v3`, see the [code repo](https://huggingface.co/datasets/legexbenchmark/code)).
Harvey and Legora take one prompt per review-table column, so the same rules
were entered per column; the platforms' workflow assistants rephrase prompts
on entry, and the as-run column prompts therefore differ in wording between
runs. The as-run prompts of every Harvey and Legora run are shipped verbatim
in `prompts/` (Harvey's 18 May set is also reproduced in the paper appendix).
All runs encode the same 14-field schema (see the
[goldensets repo](https://huggingface.co/datasets/legexbenchmark/goldensets)
for the schema definition).

## File layout

```
data/<cc>/inference_harvey.jsonl
data/<cc>/inference_harvey_2.jsonl
data/<cc>/inference_gemini.jsonl
data/<cc>/inference_gpt.jsonl
data/<cc>/inference_legora_1.jsonl
data/<cc>/inference_legora_2.jsonl
prompts/prompts_harvey.jsonl
prompts/prompts_harvey_2.jsonl
prompts/prompts_legora_1.jsonl
prompts/prompts_legora_2.jsonl
```

The `prompts/` files hold the per-column prompts as run, one record per
(run, field): `model`, `inference_date`, `field`, `prompt`.


### Record schema

Each record corresponds to one (judgment, system) pair:

| Key | Type | Description                                                                                                                                                                                                     |
|---|---|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| `case_id` | string | Identifier matching the goldenset `case_id`.                                                                                                                                                                    |
| `link` | string \| null | URL to the original judgment.                                                                                                                                                                                   |
| `legal_subject_judgement` … `defendant_no1_ISIC1_industry_category` | string \| null | The 14 extracted fields, same schema as the goldensets. All values are strings (numbers as digit strings) so every column is single-typed, this fixes the `ArrowInvalid` load failure of the first CSV release. |
| `model` | string | Model identifier emitted by the inference pipeline (see table above).                                                                                                                                           |
| `inference_date` | string | The date the producing run was executed/exported (e.g. `2026-08-01` for both Legora runs; `2026-05-18`/`2026-06-30` for Harvey), so run variants remain traceable per record.                                   |
| `error` | string \| null | Non-empty only if the model call failed for that case.                                                                                                                                                          |
| `comment` | string \| null | Human-readable note listing every value changed during cleaning; `null` when the record needed no change.                                                                                                       |
| `original_input` | string | JSON object mapping each cleaned field to its 1:1 pre-cleaning value; `"{}"` when nothing changed.                                                                                                              |

See [goldensets](https://huggingface.co/datasets/legexbenchmark/goldensets) for a description of the data columns.

## Cleaning

Long-form refusals ("…the date is not stated … N/A") and prose/citations by the model
not following instructions were removed or standardized via a two-step, human-reviewed pipeline
(`legex-refusals-scan` > review > `legex-refusals-apply`, see the
[code repo](https://huggingface.co/datasets/legexbenchmark/code)). Every
change is recorded per record in `comment` / `original_input`, so the raw
model output can always be reconstructed.

## Coverage and known limitations

- Only 15 of Spain's 129 annotated judgments have Harvey output (the ingest of
  the remaining case packets failed before the 2026-06-30 export and was not
  re-run); evaluation counts each system only on the judgments it actually
  produced output for, and reports per-system denominators.
- Hong Kong `case_id`s that were shared by two different documents (one CFA
  case number covering e.g. a substantive judgment and a costs ruling) are
  disambiguated with the source document id, e.g. `FACV12/2022-DIS154422`,
  matching the goldensets release.

- Inference ran on the full sampled case packets (up to 130 cases per
  jurisdiction), so files can contain more cases than the goldenset has
  expert-reviewed rows. Evaluation joins on `case_id` against the goldenset.
- Some Harvey exports are partial: `es` has 15 rows, `np` 108, `ph` 115,
  `hk` 116, `rs` 126, `au` 128, `de` 129, all other files have 130. The
  `harvey_2` run shares most of these gaps: `es` 15, `hk` 102, `ph` 115,
  `rs` 126, `de` 129, all other files 130.
- Legora returned empty review tables for Armenia (`am`), Georgia (`ge`), and
  most Nepalese (`np`) judgments — the rows exist but their fields are empty
  (Legora appears not to support non-Latin scripts). Evaluation counts each
  system only on the judgments it actually answered. Legora `hk` files have
  116 rows.
- For Switzerland, `gemini` and `gpt` emit `case_id`s with different
  separators than the goldenset (e.g. `4A_426/2024` vs `4A-426-2024`), join
  after normalising separators.
- A few files contain duplicate `case_id` rows.

## Loading

```python
import pandas as pd

# One (jurisdiction, system) pair
df = pd.read_json(
    "hf://datasets/legexbenchmark/inference-results/data/us/inference_gpt.jsonl",
    lines=True,
)
```

or with the `datasets` library (one config per system):

```python
from datasets import load_dataset

ds = load_dataset("legexbenchmark/inference-results", "harvey", split="switzerland")
ds = load_dataset("legexbenchmark/inference-results", "gpt", split="united_states")
ds = load_dataset("legexbenchmark/inference-results", "legora_1", split="switzerland")
```

To compare against the gold annotations, join on `case_id` after loading the
matching split from
[`legexbenchmark/goldensets`](https://huggingface.co/datasets/legexbenchmark/goldensets).


## License

MIT.