Model Card for AS-detection_gemma-2-9b-it

Gemma-2-9b-it finetuned on the GDN-CC dataset for the task of Argumentative Structure Detection. This is the best model for AS detection and the one used to annotate GDN-CC-large.

Uses

It is recommended to use it with the vLLM framework:

from vllm import LLM, SamplingParams

llm = LLM(model="LequeuISIR/AS-detection_gemma-2-9b-it", 
          max_model_len=2048)

tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b-it")

sampling_params = SamplingParams(temperature=0.2, max_tokens=2000)

messages = [
            {"role": "user", "content": f"{PROMPT}texte initial:\n {item["text"].strip()}\n\n segment à annoter:\n{item["AU"].strip()}"}
        ]
prompt_string = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True
)
outputs = llm.generate(formatted_prompts, sampling_params)

with the prompt being:

PROMPT= """
Je vais te donner un segment de texte d'opinions en français. Ton travail est de segmenter ce texte et attribuer à chaque segment un type. \
les types possibles sont CLAIM, PREMISE et SOLUTION, et UNIQUEMENT ceux-là. Ci-dessous la définition de chaque type:\n \
- SOLUTION: une proposition d'action (concrête et réalisable ou non) à prendre pour résoudre un problème.\n \
- CLAIM: l'expression d'une opinion comme affirmation, que n'apporte pas de solution mais plutôt exprime un sentiment.\n \
- PREMISE: une justification, un argument, ou un exemple qui soutient une affirmation ou une solution.\n\n \
Cette tâche est EXTRACTIVE, to dois copier le texte de chaque segment exactement comme il est écrit, incluant les majuscules et la ponctuation. \
l'intégralité du texte doit être segmenté. il n'y a pas forcément tous les types de segments, et plusieurs segments peuvent avoir le même type. \
Tu DOIS ressortir la segmentation en suivant la forme exacte de l'exemple, incluant le "-" pour chaque segment. \n\n \

- [CLAIM] Affirmation 1\n \
- [SOLUTION] Solution 1\n \
- [CLAIM] Affirmation 2\n \
- [PREMISE] argument 1\n \
...

Je vais te donner le texte initial et le segment, et tu dois sortir la liste des segments et leur types sous la forme "- [TYPE] SEGMENT", et rien d'autre.
"""

BibTeX:

@inproceedings{lequeu-etal-2026-gdn,
    title = "The {GDN}-{CC} Dataset: Automatic Corpus Clarification for {AI}-enhanced Democratic Citizen Consultations",
    author = {Lequeu, Pierre-Antoine  and
      Labat, L{\'e}o  and
      Cave, Laur{\`e}ne  and
      Lejeune, Ga{\"e}l  and
      Yvon, Fran{\c{c}}ois  and
      Piwowarski, Benjamin},
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.1523/",
    doi = "10.18653/v1/2026.acl-long.1523",
    pages = "32976--33006",
    ISBN = "979-8-89176-390-6",
    abstract = "LLMs are ubiquitous in modern NLP, and while their applicability extends to texts produced for democratic activities such as online deliberations or large-scale citizen consultations, ethical questions have been raised for their usage as analysis tools. We continue this line of research with two main goals: (a) to develop resources that can help standardize citizen contributions in public forums at the \textbf{pragmatic level}, and make them easier to use in topic modeling and political analysis; (b) to study how well this standardization can reliably be performed by small, open-weights LLMs, \textit{i.e.} models that can be run locally and transparently with limited resources. Accordingly, we introduce \textbf{Corpus Clarification} as a preprocessing framework for large-scale consultation data that transforms noisy, multi-topic contributions into structured, self-contained argumentative units ready for downstream analysis. We present \textbf{GDN-CC}, a manually-curated dataset of 1,231 contributions to the French \textit{Grand D{\'e}bat National}, comprising 2,285 argumentative units annotated for argumentative structure and manually clarified. We then show that finetuned Small Language Models match or outperform LLMs on reproducing these annotations, and measure their usability for an opinion clustering task. We finally release \textbf{GDN-CC-large}, an automatically annotated corpus of 240k contributions, the largest annotated democratic consultation dataset to date."
}