Hugging Face's logo

Privoid-ai
/
Privoid-Action-1B

Text Generation
PEFT
Safetensors
English
granite
lora
meetings
action-items
decisions
information-extraction
local-inference
conversational
Model card Files
xet
 Community
Privoid-Action-1B / docs /DATASETS.md
Privoid-ai's picture
Privoid-ai
Privoid Action-1B experimental preview (LoRA adapter, validator, API, locked test results)
b68d688 verified
preview code
|
Raw History Blame Contribute Delete
12.3 kB

Privoid Action — dataset research and provenance report

Research date: 2026-10-26. "Verified" means read from the primary source (licence file or official page) on that date. "Not verified" means only secondary sources were reachable; such data is not used.

Not legal advice: the licence terms below were read and recorded, and the commercial-use conclusions should be confirmed by whoever approves the release.

1. Conclusion first

  • No permitted English corpus labels what this task needs: action items with owner, assignment/acceptance status, deadline and evidence. Published action-item work on AMI derives labels indirectly from summaries (e.g. 101 AMI meetings / 381 action items in Liu et al. 2023, arXiv:2303.16763). The ICSI action-item annotations of Purver et al. (18 meetings) are described there as not publicly available, and inter-annotator agreement for action items has been reported as low (κ ≈ 0.36 on ICSI).
  • AMI and ICSI (both CC BY 4.0, verified) provide real meetings with speaker channels, dialogue acts and abstractive summaries linked to supporting dialogue acts. They are used for real-meeting evaluation (recall-only, "silver"), as the source of segments for human annotation, and, for ICSI, as a separate transfer domain. A summary is not treated as complete action-item ground truth.
  • Complete labels therefore come from (a) controlled synthetic templates (disclosed; see §4) and (b) human annotation of AMI segments (queue prepared: 220 segments; docs/ANNOTATION_GUIDE.md). Until (b) exists, the model card must say that no real-meeting precision, owner or deadline accuracy has been measured.

2. Candidates

2.1 AMI Meeting Corpus — manual annotations v1.6.2 (release notes say 1.7) — used

Item Finding
Source https://groups.inf.ed.ac.uk/ami/download/ → ami_public_manual_1.6.2.zip (dated 10-Apr-2017; unchanged since 16 June 2014)
Licence CC BY 4.0, verified in LICENCE.txt inside the zip and at https://groups.inf.ed.ac.uk/ami/corpus/license.shtml ("The AMI corpus and its annotations are released under … CC BY 4.0"). A CC BY 4.0 licence permits commercial use and adaptation with attribution. (openslr.org still shows an old NC-SA notice; the official site and the in-archive licence supersede it.)
Attribution Credit the AMI Consortium / University of Edinburgh, link the licence, and state changes. NOTICE text in the release folder.
Redistribution Allowed under CC BY with attribution. We do not redistribute it anyway: derived files stay local.
Download 22,887,865 bytes, SHA-256 b56e5babb2496b8795deeeda7e71178d7fbc9963f94276cf2a3f4b56ebbc9f9d. No registration or agreement. Text and annotations only; no audio or video downloaded.
Language / domain English (many non-native speakers); mostly scenario-driven product-design meetings (remote-control design team, roles PM/ID/UI/ME), plus some natural meetings (IB, IN, EN series)
Size (measured) 139 meetings with dialogue-act annotation; 66,043 turns; 682,390 words; median 16.9K tokens per meeting (p90 28K, max 47K) with the Granite tokenizer
Speakers 4 channels per meeting with roles (corpusResources/meetings.xml). Participant codes are pseudonymous; first names occur in speech.
Action / decision coverage Abstractive summaries (~200 words, one annotator) with ACTIONS (300 sentences, 256 linked) and DECISIONS (680 sentences, 624 linked) sections; dialogue-act types include Suggest, Offer, Elicit-Offer-Or-Suggestion; decision regions (decision/manual, word spans) in 48 meetings / 288 regions. No owner IDs, assignment status, acceptance or deadline fields. Summary owners are usually roles.
Evidence alignment extractive/*.summlink.xml links summary sentences to dialogue acts. The links are topical: median span 1 turn, but p90 = 125 turns.
Split Official AMI full-corpus partition by meeting series (train / dev ES2003, ES2011, IS1008, TS3004, TS3006, IB4001-4004, IB4010-4011 / test ES2004, ES2014, IS1009, TS3003, TS3007, EN2002)
Personal information Recorded with consent for research release. Contains first names and some personal small talk. Kept local; no transcript text in the release or the logs.
Use here Silver recall evaluation (dev/test), human-annotation queue, segment-length measurements, length/latency tests. Not used as training targets unless a human labels the segment.

2.2 ICSI Meeting Corpus — core + contributed annotations (ICSI_plus_NXT v1.0, 22-July-2016) — used (evaluation only)

Item Finding
Source https://groups.inf.ed.ac.uk/ami/icsi/download → ICSI_plus_NXT.zip (includes the core transcripts and dialogue acts, plus contributed annotations)
Licence CC BY 4.0, verified in LICENCE.txt inside the zip and at https://groups.inf.ed.ac.uk/ami/icsi/license.shtml ("The ICSI corpus and its annotations are released under … CC BY 4.0"). The contributions README notes that their quality is not guaranteed.
Download 55,491,177 bytes, SHA-256 e4ab79a8b638bdd29121d744c452c680fe23791e406ffe5125e2f5b118923439. No registration. Text only (audio is offered separately; not downloaded).
Language / domain English; natural weekly research-group meetings at ICSI (speech, NLP, networking), 3–10 speakers
Size (measured) 75 meetings, 64,252 turns, 771,477 words; median 31K tokens per meeting (p90 49K, max 88K). One meeting (Bed017) has a single turn over 1,536 tokens and is excluded from segmentation (recorded in skipped_meetings.json).
Speakers Participant codes (me011, fe004, …); no names in metadata
Coverage MRDA-style dialogue acts; abstractive summaries with ABSTRACT, DECISIONS (281 sentences, 264 linked), PROBLEMS and PROGRESS. No actions section, no owners or deadlines.
Evidence alignment summlink for 59 meetings (decisions → dialogue acts)
Personal information As for AMI; some personal chatter, participant codes only
Use here Transfer-domain test (never trained on): silver decision recall, output validity and prediction rates on a different meeting genre

2.3 QMSum (Yale-LILY) — not downloaded

Query-based summaries over 232 meetings (Academic = ICSI, Product = AMI, Committee = Welsh Parliament / Parliament of Canada committees), with relevant_text_span alignments. The repository states an MIT licence (repo page; not re-checked against every upstream source's terms). It adds no action, owner, deadline or acceptance labels beyond the AMI/ICSI material we already have. The committee transcripts could serve as a further transfer domain in a later release, after their own (parliamentary) licence terms are checked. Source: https://github.com/Yale-LILY/QMSum

2.4 MeetingBank (Hu et al., ACL 2023) — excluded

City-council meetings (6 US cities, 1,366 meetings, 3,579 hours) with summaries and minutes. The Hugging Face dataset cards (huuuyeah/MeetingBank_Audio, lytang/MeetingBank-transcript) state CC BY-NC-SA 4.0; the Zenodo record's terms were not verified. Non-commercial, so excluded from training. It would be acceptable only for internal non-commercial evaluation, after the original terms are confirmed. Public-meeting recordings contain names of officials and members of the public.

2.5 ELITR Minuting Corpus (UFAL, Charles University) — not used (licence not verified)

Czech and English meeting transcripts (120 English, computer-science domain) with manually written minutes and transcript–minute alignments; used in AutoMin. The LINDAT catalogue page (http://hdl.handle.net/11234/1-4692) could not be loaded on 2026-10-26, so the licence and access conditions could not be confirmed. Sections are censored for privacy and audio is not released. Potentially valuable for minutes alignment; revisit after confirming the licence.

2.6 Others reviewed

  • AMC-A (Alibaba, arXiv:2303.16763): 424 Chinese meetings with sentence-level action-item labels. Out of scope for an English first release; licence not checked.
  • MeeQA (arXiv:2305.08502) and MISeD (arXiv:2405.01121): question answering and information-seeking dialogues over AMI/ICSI/QMSum meetings. No action-item labels; licences not verified. Not used.
  • SAMSum, DialogSum: chat or dialogue summarisation, not meetings, with non-commercial licences (SAMSum). Not used.

3. Files actually downloaded and prepared (authoring run, 2026-10-26)

File Bytes SHA-256
data/raw/ami/ami_public_manual_1.6.2.zip 22,887,865 b56e5bab…f9d
data/raw/icsi/ICSI_plus_NXT.zip 55,491,177 e4ab79a8…439

Prepared outputs, all reproducible from the zips by notebook sections 4–6 (exact record: authoring/RUN_authoring_data_prep.executed.ipynb):

  • data/prepared/corpus-v1-seg1536-ov8/: per-split segment files with silver hints, manifest.json (SHA-256 per file), split_manifest.json (meeting IDs per split), skipped_meetings.json, corpus_stats.json.
  • data/prepared/synth-1.0.0-smoke/: smoke-size synthetic set only (the full set is generated by you), with manifest.json, dedup_report.json, quarantine.jsonl and the label-review CSV.
  • data/annotations/ami_annotation_queue_*.jsonl: 220 AMI segments (train/dev/test series) for human labelling.

Label provenance per example: label_origin ∈ {synthetic_template, human, corpus_summary_link, none}; label_completeness ∈ {complete, partial}.

4. Augmentation strategy (synthetic data), and why

Gap addressed: the commitment-level distinctions the product needs, which no permitted corpus labels: suggestions vs commitments, requests with and without acceptance, acceptance in a later turn, multiple actions in one turn, shared and missing owners, non-participant owners, relative and absolute deadlines, corrections and withdrawn commitments, conditional tasks, decisions with and without actions, options, quoted historical commitments, interruptions and disfluency, ASR-like errors, similar speaker names, and embedded instructions.

How: a deterministic local template generator (privoid_action/synth.py, version 1.0.0). It uses no paid API and no external model; an optional paraphrase pass with the local Granite model rewrites only non-evidence filler turns. Gold labels are built together with the text, re-validated with the same strict contract as model output, and quarantined if they fail. Expected deadline normalisation is specified per phrase, independently of the normaliser (0 mismatches over 4,865 generated examples during authoring). Exact and near-duplicate removal (MinHash, Jaccard ≥ 0.85) keeps the evaluation copy. Every phenomenon is sampled into a human label-review CSV, and full training requires a recorded sign-off.

Disclosure: synthetic proportion of training data is recorded per run (100% unless human AMI labels are added). Results on template data (test_iid, test_heldout) are reported separately and are never presented as real-meeting performance. During authoring, a simple rules baseline already reached action F1 ≈ 0.81 on synthetic dev (52 examples) but recalled only 1 of 16 summary-linked actions (0.06) in a 30-segment real AMI dev sample. That gap is why the human-labelled AMI test matters.

5. Sources