--- pretty_name: UN Human Rights Voting Records (CHR 1946–2006 · HRC 2006–) homepage: https://lszoszk.github.io/hrc-voting/ license: other license_name: polyform-noncommercial-1.0.0 license_link: https://polyformproject.org/licenses/noncommercial/1.0.0/ annotations_creators: - expert-generated - found language_creators: - found multilinguality: - monolingual language: - en source_datasets: - original task_categories: - tabular-classification - text-classification - text-retrieval task_ids: - multi-class-classification size_categories: - 10K. ## Configs Three fact tables and two dimension tables. | config | rows | one row is | period | |---|---|---|---| | `resolutions` *(default)* | 6,346 | a catalogued resolution, decision or amendment | 1946–2026 | | `votes` | 80,159 | one country's position on one resolution | 1947–2026 | | `clauses` | 96,451 | one preambular / operative / annex clause | 1993–2026 | | `countries` | 154 | a state that cast at least one roll-call vote | — | | `subjects` | 1,033 | one OHCHR controlled subject heading | — | ```python from datasets import load_dataset res = load_dataset("lszoszk/hrc-voting", "resolutions", split="train") votes = load_dataset("lszoszk/hrc-voting", "votes", split="train") clauses = load_dataset("lszoszk/hrc-voting", "clauses", split="train") countries = load_dataset("lszoszk/hrc-voting", "countries", split="train") subjects = load_dataset("lszoszk/hrc-voting", "subjects", split="train") ``` Join keys: `record_id` links the three fact tables; `iso3` links `votes` to `countries`; `subject` links `resolutions` / `votes` / `clauses` to `subjects`. The two dimension tables exist so that caveats which would otherwise be prose are machine-readable — which states no longer exist, which ISO code spans a change of representation rather than of state, and which subject headings name a country situation rather than a theme. ## Read this before analysing These are the four things most likely to produce a wrong result. The first two are mistakes this project made and corrected; derived columns are provided so you do not have to repeat them. **1. Only 27% of resolutions have a per-country roll-call.** 1,705 of 6,346. The rest were adopted without a vote or published as totals only, so no individual positions exist. Roll-call statistics describe *contested* politics, not the Council's output — a well-documented selection problem (Fjelstul, Hug & Kilby 2025). Filter on `has_rollcall`. **2. An abstention is not a vote against.** Under the chamber's own rules a member that abstains is "considered as not voting", so the outcome turns on Yes vs No alone. Taking `max(yes, no, abstain)` as the winning side — the obvious-looking move — scores the winning Yes bloc as *defeated* on **53 adopted resolutions** in this corpus, where more members abstained than voted Yes but the text still carried. Use `prevailing_side` / `adopted`, which are computed correctly (Yes vs No, null on a 6-case tie). **3. Amendments are not resolutions.** 731 catalogued items are amendments to draft texts, and 462 of them went to a recorded vote — 27% of all roll-calls, almost all post-2010. A Yes on a hostile amendment frequently *opposes* the parent text's aim, so pooling them with resolutions inverts meaning. Filter on `is_amendment`. **4. ~9% of roll-calls do not reconcile with the official totals.** For 148 recorded votes the sum of individual Yes/No/Abstain positions does not match the resolution's official 996 totals, almost always by a small margin in abstentions or non-participation. These are inconsistencies in OHCHR's own cataloguing and are preserved as-is; the official totals are treated as authoritative. Filter on `rollcall_reconciles`. ## Config: `resolutions` One row per catalogued item. Vote totals are the **official** figures from MARC 996, which exist for many resolutions that have no per-country breakdown. ```json { "record_id": "35356", "symbol": "A/HRC/RES/61/10", "title": "Torture and other cruel, inhuman or degrading treatment or punishment: mandate of the Special Rapporteur : resolution /", "date": "2026-03-30", "year": 2026, "body": "HRC", "vote_type": "ADOPTED WITHOUT VOTE", "adoption_mode": "adopted without a vote", "is_amendment": false, "subject": "TORTURE AND INHUMAN TREATMENT - SPECIAL RAPPORTEURS", "agenda_item_no": "3.", "agenda_item_title": "Promotion and protection of all human rights, civil, political, economic, social and cultural rights, including the right to development.", "main_sponsors": "Denmark", "meeting": "Adopted at the 52nd meeting", "yes": null, "no": null, "abstain": null, "nonvoting": null, "total": null, "n_rollcall": 0, "has_rollcall": false, "prevailing_side": null, "adopted": null, "rollcall_reconciles": null, "url_resolution": "http://undocs.org/A/HRC/RES/61/10", "url_draft": "http://undocs.org/A/HRC/61/L.10", "record_url": "https://searchlibrary.ohchr.org/record/35356" } ``` - `record_id` — stable Invenio id (MARC 001); join key - `symbol` — e.g. `A/HRC/RES/61/29`, `E/CN.4/RES/2005/1` - `title`, `date`, `year`, `body` — `CHR` or `HRC` - `vote_type` — raw catalogue value (MARC 591); `adoption_mode` — readable form - `is_amendment` — an amendment to a draft text, detected from two agreeing catalogue signals (a `: amendment /` title marker and the MARC 993 draft note) - `subject` — OHCHR's own controlled subject heading (MARC 991$d); **not** ML-derived - `agenda_item_no`, `agenda_item_title`, `main_sponsors`, `meeting` - `yes`, `no`, `abstain`, `nonvoting`, `total` — official totals (MARC 996), null when the catalogue records none - `n_rollcall`, `has_rollcall` — number of per-country rows in `votes` - `prevailing_side` — `Y`, `N`, or null (tie / no totals) — **see caveat 2** - `adopted` — `prevailing_side == "Y"`, null when undetermined - `rollcall_reconciles` — whether the roll-call tally matches the official totals; null when there is no roll-call — **see caveat 4** - `subject_is_country_situation` — whether the subject heading names a state or territory rather than a theme; null only where the record carries no subject (219 rows). See the `subjects` config for how this is derived - `url_resolution`, `url_draft`, `record_url` Adoption modes: adopted without a vote 3,835 · recorded vote 1,713 · non-recorded vote 371 · withdrawn 236 · non-recorded unanimous 142 · not considered 33 · non-recorded no information 15 · recorded at a closed meeting 1. Bodies: CHR 3,082 · HRC 3,264. ## Config: `votes` One row per (resolution, country). 154 states appear. - `record_id`, `symbol`, `date`, `year`, `body` - `iso3`, `country` — ISO-3166 alpha-3 and the name as catalogued - `un_regional_group` — WEOG / Eastern Europe / GRULAC / African / Asia-Pacific, assigned by present-day membership and applied across the whole period - `vote` — raw code, `vote_label` — readable - `is_cast_vote` — true for `Y`/`N`/`A` only - `prevailing_side`, `with_prevailing_side` — whether this state was on the side that prevailed; null when it cast no vote or the resolution had no Yes/No outcome - `subject`, `is_amendment` — denormalised from `resolutions` for convenient filtering Vote codes: `Y` yes 44,536 · `N` no 18,368 · `A` abstain 15,911 · `.` absent or not participating 701 · empty 643 (not a member at the time, or no position recorded). ## Config: `clauses` Clause-segmented full text of 4,437 resolutions, **1993–2026** — the earliest the OHCHR/ODS document mirrors reach, so this config covers a much shorter period than the other two. Extracted from `.doc`/`.pdf` originals and segmented on structural markers. - `record_id`, `symbol`, `year`, `body`, `subject`, `title`, `is_amendment`, `adoption_mode`, `adopted` - `clause_index`, `clause_label` — e.g. `PP3`, `OP7`, `OP2(a)`, `AX1` - `clause_type` — `preambular` 35,670 · `operative` 42,957 · `operative_subitem` 12,498 · `annex` 2,082 · `other` 3,244 - `text`, `n_chars`, `n_words` Four **experimental** columns score top-level operative clauses only (null elsewhere). Each operative paragraph opens by drafting convention with a controlled verb, matched against UNITAR's *Terms Used in Resolutions* glossaries; 98.5% of operative clauses resolve to a verb. - `operative_verb` — canonical form, e.g. `Requests`, `Calls upon`, `Condemns` - `directive_force` — 5 decisional (*Decides, Demands, Establishes*) · 4 strong exhortation (*Urges, Calls upon*) · 3 tasking (*Requests, Recommends*) · 2 invitation (*Invites, Encourages*); null for non-directive verbs - `sentiment` — −2 condemnation … 0 neutral registering … +2 commendation; null for directive verbs and for verbs carrying no valence code. For *Notes*, *Takes note*, *Recognizes* and *Expresses* the score is resolved **per clause** from the object ("*with concern*" → −1, "*with appreciation*" → +1), following a 2001 UN Office of Legal Affairs opinion (endorsed in GA decision 55/488) that the bare terms are neutral - `creates_or_tasks_machinery` — the delegation axis: true only when a *directive* clause establishes, renews or extends a mechanism, or names a UN actor and gives it something to do (17.4% of operative clauses). A passing mention of a rapporteur or a report does **not** count; scoring mentions instead put 40% of clauses on this axis and imported a spurious trend, since OHCHR and the special procedures only expanded after 1993 **These columns measure rhetorical / directive hardness, not legal bindingness.** Even "Decides" binds only under Security-Council Chapter VII; CHR and HRC resolutions are formally non-binding whatever the verb. Tagging only the leading verb also omits the addressee and the precision of the ask. The tiers are grounded in the "legalization" framework (Abbott, Keohane, Moravcsik, Slaughter & Snidal 2000) and a content-analysis approach developed for Security Council resolutions (Benson & Tucker 2022); the exact rank within a tier and the numeric spacing are an author-defined coding. Articles of annexed declarations and protocols are labelled `annex`, not `operative` — they are treaty-style text, not commitments of the organ, and conflating them shifts every operative-verb statistic. `subject_is_country_situation` is carried here too. ## Config: `countries` Dimension table, 154 states. Join on `iso3`. - `iso3`, `name` — display name; the catalogue's own spellings vary over time and are preserved in `votes.country` - `un_regional_group`, `un_regional_group_code` — **applied anachronistically**: groups are present-day and the Commission's membership changed over 60 years - `n_rollcall_cells`, `n_yes`, `n_no`, `n_abstain`, `n_absent`, `n_no_position` - `first_vote_year`, `last_vote_year` — first and last **recorded vote**, not membership. 12 calendar years contain no recorded vote at all, so a gap here is not evidence a state was off the Council - `is_historical_state` — true for 5 states that no longer exist: `CSK` Czechoslovakia, `SUN` Soviet Union, `YUG` Yugoslavia, `DDR` and `GER` the two Germanys - `representation_break_year`, `representation_break_note` — set only for `CHN`: the China seat passed from the Republic of China to the PRC in October 1971 and the catalogue records both under one code, so a continuous 1947–2026 China series is not one continuous actor ## Config: `subjects` Dimension table, all 1,033 controlled subject headings (MARC 991$d) — OHCHR's own cataloguing vocabulary, **not** ML-derived, so every label is traceable to the source. Join on `subject`. - `subject` — the heading as catalogued, upper-case - `is_country_situation` — 307 of 1,033 name a state or territory - `n_resolutions_all`, `n_resolutions_recorded` — how many resolutions carry it, in the full catalogue and in the roll-call subset The thematic / country-situation split is a **name-matching heuristic**: catalogued state names, plus an explicit list of territories and non-member states that never sat on the Council and so never appear in the vote data (Golan Heights, Western Sahara, Kosovo, Darfur, Xizang, …). It is imperfect at the margins. The same matcher produces the dashboard's Topics filter, so the two cannot disagree. ## Known data-quality notes - **Three MARC `967$b` codes contradicted their country name** and are repaired by the pipeline, which trusts the name: rec 21027 HUNGARY was coded `HND`, rec 21028 NORWAY was `PER`, rec 30508 BRAZIL was `BEN`. Each had been moving one vote onto another state's profile. - **One roll-call row has no country** (rec 24737, `E/CN.4/RES/6(XX)`, 1964) — a malformed MARC field. It is kept for fidelity to the source, with `iso3` null. - **One vote-code typo** (`y` → `Y`, UK 2011) is normalised. Nothing else is altered. - **12 calendar years contain no recorded vote at all** (1949, 1951, 1953, 1956–57, 1959–63, 1965, 1973). A state's absence from the roll-call in those years says nothing about its membership. - **`CHN` spans a change of representation, not of state** — the China seat passed from the Republic of China to the PRC in October 1971; the catalogue records both as `CHN`. - **`un_regional_group` is applied anachronistically.** Groups are present-day; the Commission's membership and the groups themselves changed over 60 years. - Historical states appear under their own codes (`CSK` Czechoslovakia, `SUN` Soviet Union, `YUG` Yugoslavia, `DDR`/`GER` the two Germanys). ## Suitable for Roll-call and ideal-point analysis of a small elected chamber (as opposed to the UN General Assembly, which most of the quantitative literature studies); bloc and alignment work; agenda and subject analysis using the catalogue's own controlled vocabulary; clause-level retrieval over UN human-rights text; and study of drafting language and institutional commitment. Most quantitative UN-voting findings come from the **General Assembly** — universal membership, plenary voting. The Commission (53 members elected by ECOSOC) and the Council (47, elected by the GA under regional quotas) are small elected chambers where membership itself is selected and absence is marginal. Treat GA findings as hypotheses here, not established facts. ## Provenance and reproduction Source: OHCHR Search Library, [Voting collection](https://searchlibrary.ohchr.org/search?c=Voting). The full pipeline — harvest, parse, build — is in the repository and this package is rebuilt from committed outputs by [`scripts/prepare_hf_dataset.py`](https://github.com/lszoszk/hrc-voting/blob/main/scripts/prepare_hf_dataset.py). Methods and caveats are documented in the dashboard's Methodology tab and in `notes/AUDIT-2026-07.md`. ## Licence Code and data are released under the **PolyForm Noncommercial License 1.0.0**: research, education, non-profit and personal use are permitted; **commercial use requires a separate licence from the author.** The underlying voting records are public UN documents from the OHCHR Search Library. ## Citation > Szoszkiewicz, Ł. (2026). *OHCHR Voting Records — CHR · HRC* (v1.0.0) [Software]. > Zenodo. https://doi.org/10.5281/zenodo.21281232 ```bibtex @software{szoszkiewicz_hrc_voting_2026, author = {Szoszkiewicz, {\L}ukasz}, title = {OHCHR Voting Records --- CHR {\textperiodcentered} HRC}, year = {2026}, version = {1.0.0}, publisher = {Zenodo}, doi = {10.5281/zenodo.21281232}, url = {https://lszoszk.github.io/hrc-voting/} } ``` Built and maintained by [Łukasz Szoszkiewicz](https://lszoszk.github.io/) (Adam Mickiewicz University, Poznań) · [ORCID 0000-0001-6671-2893](https://orcid.org/0000-0001-6671-2893).