nashicloud/lynceus-silver

Lynceus silver

Version française : README.fr.md

The training set of Project Lynceus, which describes the rhetorical techniques a web page uses without judging its authors. 2,053 French and English web pages, each read by a panel of three open-weight language models following the Lynceus annotation guide.

No page text is included, not even excerpts. Techniques are located by character offsets; reconstruire.py fetches the pages, checks they are unchanged, and rebuilds the excerpts. For training only, never for evaluation. Publishers and people named in a page can ask for its removal (see below).

What the dataset contains

silver: 2,053 pages, one line per page.

Field	Content
url, domaine, titre, langue	the page (French fr or English en)
content_hash	SHA-256 of the extracted text, whitespace collapsed: tells whether the page changed since capture
categorie, categories_acceptables	the nature of the page, one of ten categories (below)
grade	one or two letters from A to E: the quality of the argumentation
techniques	the techniques found, among the 31 of the Lynceus taxonomy (below), each with its accord (a+b: seen by both annotators; arbitre: seen by one and confirmed by the arbiter), the position of its excerpt (debut, fin, in characters of the normalized page text) and its fingerprint (empreinte_extrait). The excerpt text itself is not published
panel, guide, prompts, decision	provenance: which models, which guide version, consensus or arbitration
lecture_a, lecture_b	the two original readings, reduced to their category and techniques
intention_recherche	the category the search query was aiming at; not a label

Field names and labels are those of the Lynceus taxonomy, in French.

• Categories: information (news), opinion, analyse_expertise (analysis, expertise), satire, publicite_sponsorise (advertising, sponsored), temoignage (personal testimony), contenu_confessionnel (religious content), pseudo_science, theorie_du_complot (conspiracy theory), autre (other).
• Techniques (31): appeal to fear (appel_a_la_peur), artificial urgency (urgence_artificielle), appeal to outrage (appel_a_l_indignation), hidden truth (verite_cachee), us versus them (eux_contre_nous), hyper-intentionality (hyper_intentionnalisme), Gish gallop (millefeuille_argumentatif), persecuted hero (heros_persecute), cherry picking, correlation as causation (correlation_causation), false dilemma (faux_dilemme), slippery slope (pente_glissante), hasty generalization (generalisation_abusive), anecdotal evidence (preuve_anecdotique), appeal to nature (appel_a_la_nature), pseudo-scientific jargon (jargon_pseudo_scientifique), straw man (epouvantail), whataboutism (whataboutisme), moving the goalposts (deplacement_des_buts), just asking questions (je_pose_des_questions), impossible expectations (attentes_impossibles), anonymous authority (autorite_anonyme), fake expert (faux_expert), missing sources (absence_de_sources), circular sources (sources_circulaires), false balance (faux_equilibre), ad hominem, attributing bad intent (proces_d_intention), dehumanization (deshumanisation), commercial conflict of interest (conflit_interet_commercial), miracle cure (solution_miracle). Definitions: TAXONOMIE.md.

banc: 117 pages read by a human without seeing the panel's reading, then confronted with it. Each page carries the blind reading (aveugle) and the reference corrected by the confrontation verdicts (corrigee). It is a development bench to compare models. It is not the Lynceus test set, which stays human, blind, and outside this repository.

Only the blind reading is independent of the panel. The corrected reference was established after seeing the panel's reading: it favours the panel, and any reading that resembles it. Look at both references when comparing models.

Page text is not included

It belongs to its publishers, and the dataset contains none of it, not even excerpts. reconstruire.py downloads each page while respecting robots.txt, extracts it with trafilatura 2.2.0 as at capture time, and compares its fingerprint: only pages marked identique surely carry the dataset's annotations. For those, it rebuilds the excerpt of each technique from its position and checks it against its fingerprint. A page changed or removed since capture is reported, never corrected. On 60 pages drawn at random in September 2026, 56 were identical (93 %), and all 72 excerpts of those pages were rebuilt and verified.

pip install httpx trafilatura==2.2.0
python reconstruire.py data/silver.jsonl --sortie pages.jsonl --contact you@example.org

How the dataset was made

1. Discovery through the Brave Search API, following a versioned query plan that searches for topics, never for verdicts. Pages of the Lynceus corpus are excluded. At most 10 pages per website.
2. Capture by a crawler that identifies itself, respects robots.txt and waits 5 seconds between two pages of the same site. Texts of 200 to 60,000 characters, French or English.
3. Double annotation: two models from different providers read each page without seeing each other. Every excerpt is checked word for word against the page; an excerpt not found is rejected.
4. Arbitration by a third model, only on disagreement. It settles the category and the grade, and confirms or discards passages flagged by a single annotator. It cannot add anything.

Main panel: GLM-5.3 (z-ai/glm-5.3, Z.AI, low reasoning effort), DeepSeek V4 Pro (deepseek/deepseek-v4-pro), MiniMax M3 (minimax/minimax-m3), all called through OpenRouter at temperature 0. Kimi K3 (moonshotai/kimi-k3) arbitrated the pilot and part of the first batch. The annotation guide evolved from version 1.0 to 1.4; each page records the version it was read under.

Quality

Measured on the 117 pages of the bench:

Against	Category	Techniques: precision	recall
the corrected reference	0.96	0.92	0.83
the blind human reading	0.83	0.40	0.62

The true quality lies between the two. The confrontation is not blind and flatters the panel; a single reader's blind reading misses techniques (their own recall against the corrected reference is 0.4 to 0.6).

• The category is reliable. Cohen's kappa between the two annotators is 0.62 to 0.83 depending on the batch; under guide 1.4, on 10 hard pages chosen on purpose, the category was right 10 times out of 10 and no technique was flagged wrongly.
• Techniques present in a page are good, with high precision.
• Exact passages are only weak supervision: the two annotators often pick different sentences for the same technique.

Intended use, and what not to do with it

• To train a model that spots rhetorical techniques (encoder, small generative model). Weight techniques by accord.
• Never to evaluate a model: annotations written by models would make the measure circular.
• Not to judge a source. Lynceus describes techniques in a page, not the reliability of an outlet. intention_recherche and the composition of the dataset reflect a search plan, not a reputation.

Known limitations

• Three models from three families, but all language models: a bias they share escapes voting and arbitration.
• A single human reader for the bench.
• The composition follows the search plan, not a random sample of the web; rare categories are deliberately over-represented.
• Some pages have changed or disappeared since capture.

Right of reply and removal

Lynceus promises a right of reply for any content it analyses. The publisher of a page, or a person named in it, can ask for its removal from the dataset: open a discussion on this repository, or an issue on Project-Lynceus, giving the page address. The page is removed in the next version of the dataset, no justification needed. A reply on the substance (a disputed technique) is welcome through the same channels.

Personal data, rights, licences

• No page text is published, not even excerpts: a technique aimed at a named person (ad_hominem, proces_d_intention) appears only as a position in the page.
• Testimony pages (extraits_retires: true) are often personal accounts (health, grief, family). They do not even carry positions, only their labels.
• Text and data mining: pages were captured respecting robots.txt with an identified crawler, in line with the text and data mining exception of European Union law, which yields to a site's opt-out.
• Licence: CC BY-SA 4.0, covering the annotations (labels, positions, provenance) and nothing else. Pages remain their authors'.
• Licences of the panel models, read on their Hugging Face repositories in September 2026: GLM-5.3 and Kimi K3 under a modified MIT licence (conditions reserved for very large commercial services); DeepSeek V4 Pro under MIT; MiniMax M3 under the MiniMax Community License ("Built with MiniMax M3" notice and notification to MiniMax for commercial use). None expressly restricts redistributing outputs or using them to train other models; they cover the weights and their derivatives.