ndpvt-web/compar-ia-french-governments

compar:IA — Multilingual LLM Arena & Preference Data Collection Pipeline

This skill enables Claude to design and implement LLM evaluation arenas that collect human preference data for non-English languages, following the architecture pioneered by the French government's compar:IA platform. The core technique is a blind pairwise comparison interface backed by a FastAPI + SvelteKit stack, where users submit unconstrained prompts, receive anonymized side-by-side model responses, and cast preference votes that feed directly into RLHF/DPO training pipelines. The platform uses Bradley-Terry ranking models to aggregate pairwise judgments into a leaderboard, and applies conservative full-conversation exclusion (rather than span-level redaction) for privacy filtering.

When to Use

• When the user asks to build an LLM arena or chatbot comparison platform for any language
• When designing a pipeline to collect human preference data for DPO or RLHF fine-tuning
• When the user needs a Bradley-Terry ranking system for pairwise model evaluation
• When building a privacy-preserving data collection platform that filters PII at the conversation level
• When the user wants to create a multilingual model leaderboard from crowdsourced votes
• When deploying a blind A/B testing system for comparing LLM outputs
• When building infrastructure to route prompts to multiple inference providers (OpenRouter, HuggingFace Inference) simultaneously
• When the user needs to measure and display environmental impact of LLM inference

Key Technique

Blind Pairwise Comparison with Two-Level Feedback. The compar:IA approach collects preference data through a three-step flow: (1) the user submits an unconstrained free-form prompt, (2) two randomly selected anonymous models generate side-by-side responses, and (3) the user provides feedback at two granularities — message-level reactions (per-turn thumbs up/down) and conversation-level votes (preferred model A, model B, or tie). Only after voting does the platform reveal model identities and metadata. This dual-granularity feedback yields richer training signal than single-vote systems: conversation-level votes produce standard DPO preference pairs, while message-level reactions enable turn-level reward modeling.

Bradley-Terry Ranking from Noisy Crowdsourced Data. Rather than raw win-rate tallies, the leaderboard uses the Bradley-Terry statistical model to aggregate pairwise votes into consistent global rankings. This handles the transitivity problem (model A beats B, B beats C, but C beats A) and accounts for uneven matchup frequencies. The method treats each vote as evidence for a latent "strength" parameter per model, then fits via maximum likelihood. This is the same approach used by Chatbot Arena (LMSYS) but applied here to French-language evaluation with 250,000+ votes across 104 models.

Conservative Privacy Filtering via Full-Conversation Exclusion. Instead of attempting span-level PII anonymization (which risks incomplete masking and semantic distortion), compar:IA uses an LLM-based classifier to flag conversations likely containing personal data, then excludes the entire conversation and all associated votes. This removes approximately 5% of data but avoids the residual re-identification risks inherent in token-level redaction. This is a key architectural decision: accept modest data loss for strong privacy guarantees.

Step-by-Step Workflow

1. Define the Data Schema

Create three complementary datasets with clear separation of concerns:

# conversations schema
conversation = {
    "conversation_id": "uuid",
    "timestamp": "ISO-8601",
    "language": "fr",  # detected via langdetect or fasttext
    "turns": [
        {
            "role": "user",
            "content": "Explique-moi la photosynthèse",
            "turn_index": 0
        },
        {
            "role": "assistant_a",
            "content": "La photosynthèse est le processus...",
            "model_id": "hidden_until_vote",
            "turn_index": 1
        },
        {
            "role": "assistant_b",
            "content": "C'est un mécanisme biologique...",
            "model_id": "hidden_until_vote",
            "turn_index": 1
        }
    ]
}

# votes schema (conversation-level preference)
vote = {
    "vote_id": "uuid",
    "conversation_id": "uuid",
    "winner": "assistant_a" | "assistant_b" | "tie",
    "timestamp": "ISO-8601"
}

# reactions schema (message-level feedback)
reaction = {
    "reaction_id": "uuid",
    "conversation_id": "uuid",
    "turn_index": 1,
    "target": "assistant_a" | "assistant_b",
    "value": "positive" | "negative",
    "timestamp": "ISO-8601"
}

2. Build the Backend API (FastAPI)

Implement these core endpoints:

• POST /conversations — accepts a user prompt, selects two models via weighted random sampling, fans out inference requests in parallel, returns anonymized responses
• POST /conversations/{id}/votes — records conversation-level preference (A, B, or tie) and reveals model identities
• POST /conversations/{id}/reactions — records message-level thumbs up/down per turn
• GET /leaderboard — returns current Bradley-Terry rankings

Route inference through a provider abstraction layer that supports OpenRouter, HuggingFace Inference Providers, and direct API calls, so models can be added or swapped without code changes.

3. Implement Blind Comparison Frontend (SvelteKit or React)

Build a split-pane chat interface where:

• Left and right panels show "Model A" and "Model B" with no identifying information
• Both panels stream responses simultaneously from the same user prompt
• Multi-turn conversation is supported — users can follow up before voting
• Vote buttons (A wins / B wins / Tie) appear after at least one exchange
• After voting, a reveal overlay shows model names, parameter counts, and inference cost

4. Implement Model Selection and Pairing Strategy

import random

def select_model_pair(models: list, elo_ratings: dict) -> tuple:
    """Select two models for comparison, favoring similar-strength pairings
    to maximize information gain for Bradley-Terry fitting."""
    # Weight pairings by proximity in current ELO ratings
    # to get more discriminative comparisons
    weights = []
    pairs = []
    for i, m1 in enumerate(models):
        for m2 in models[i+1:]:
            diff = abs(elo_ratings.get(m1, 1200) - elo_ratings.get(m2, 1200))
            weights.append(1.0 / (1.0 + diff / 400.0))
            pairs.append((m1, m2))
    chosen = random.choices(pairs, weights=weights, k=1)[0]
    # Randomize left/right assignment to prevent position bias
    return chosen if random.random() > 0.5 else (chosen[1], chosen[0])

5. Implement Bradley-Terry Ranking

import numpy as np
from scipy.optimize import minimize

def fit_bradley_terry(matchups: list[dict], model_names: list[str]) -> dict:
    """Fit Bradley-Terry model from pairwise votes.
    matchups: [{"winner": "modelA", "loser": "modelB"}, ...]
    Returns dict of model_name -> strength score.
    """
    n = len(model_names)
    idx = {name: i for i, name in enumerate(model_names)}

    def neg_log_likelihood(params):
        nll = 0.0
        for m in matchups:
            wi = idx[m["winner"]]
            li = idx[m["loser"]]
            nll -= params[wi] - np.logaddexp(params[wi], params[li])
        # L2 regularization to prevent divergence
        nll += 0.01 * np.sum(params ** 2)
        return nll

    result = minimize(neg_log_likelihood, np.zeros(n), method="L-BFGS-B")
    strengths = result.x
    # Convert to ELO-scale ratings (centered at 1200)
    elo = 1200 + 400 * (strengths - strengths.mean())
    return {name: float(elo[i]) for i, name in enumerate(model_names)}

6. Build the Privacy Filtering Pipeline

Run an LLM-based PII classifier on every conversation before dataset export. Use full-conversation exclusion, not token-level redaction:

async def filter_conversation(conv: dict, classifier_model: str) -> bool:
    """Returns True if conversation is safe to publish, False if it
    likely contains PII and should be excluded entirely."""
    full_text = " ".join(t["content"] for t in conv["turns"])
    prompt = (
        "Does the following conversation contain personal information "
        "(names, addresses, phone numbers, emails, ID numbers, medical "
        "details, or any data that could identify a real person)? "
        "Answer ONLY 'yes' or 'no'.\n\n" + full_text
    )
    response = await call_llm(classifier_model, prompt)
    return response.strip().lower() == "no"

Expect approximately 5% exclusion rate. Provide a public reporting form so users can flag conversations that slipped through or were incorrectly removed.

7. Add Language Detection and Categorization

Classify each prompt by language (using fasttext lid.176.bin) and topic category. The compar:IA taxonomy uses 15 categories: Natural Science & Technology (17.7%), Education (14.3%), Business & Economics (10.1%), and others. Apply classification post-hoc for dataset analysis — do not constrain user input.

8. Integrate Environmental Impact Tracking

Use the ecologits Python library to estimate energy consumption per inference call based on token count, model architecture, and GPU manufacturing amortization. Display per-conversation carbon estimates to users after they vote.

9. Export DPO-Ready Training Data

Convert conversation-level votes into preference pairs formatted for DPO training:

def export_dpo_pairs(conversations: list, votes: list) -> list[dict]:
    """Convert arena votes to DPO training format."""
    pairs = []
    for vote in votes:
        if vote["winner"] == "tie":
            continue  # ties are excluded from DPO pairs
        conv = get_conversation(vote["conversation_id"], conversations)
        prompt = extract_user_turns(conv)
        chosen = extract_assistant_turns(conv, vote["winner"])
        rejected = extract_assistant_turns(
            conv, "assistant_b" if vote["winner"] == "assistant_a" else "assistant_a"
        )
        pairs.append({
            "prompt": prompt,
            "chosen": chosen,
            "rejected": rejected,
            "language": conv["language"]
        })
    return pairs

10. Publish Datasets Under Open License

Release three dataset splits (conversations, votes, reactions) on HuggingFace with Etalab 2.0 or CC-BY-4.0 licensing. Gate raw data for research-only access. Add a restriction that proprietary model responses may be used for analysis and evaluation but not for training.

Examples

Example 1: Building a French LLM Arena

User: "I want to build a platform like Chatbot Arena but focused on French. Help me set up the backend."

Approach:

1. Scaffold a FastAPI project with endpoints for /conversations, /votes, /reactions, and /leaderboard
2. Create a provider abstraction layer supporting OpenRouter and HuggingFace Inference APIs
3. Implement the blind model pairing logic with position randomization
4. Set up a PostgreSQL schema with the three-table design (conversations, votes, reactions)
5. Add the Bradley-Terry ranking computation as a background job that recomputes after every 100 new votes
6. Implement the PII filtering pipeline as an async post-processing step before dataset export

Output:

project/
  app/
    main.py              # FastAPI app with CORS, routes
    routers/
      conversations.py   # prompt submission, parallel inference
      votes.py           # preference recording, model reveal
      leaderboard.py     # Bradley-Terry rankings
    services/
      inference.py       # provider abstraction (OpenRouter, HF, direct)
      pairing.py         # model selection with ELO-proximity weighting
      ranking.py         # Bradley-Terry MLE fitting
      privacy.py         # LLM-based PII filter, full-conversation exclusion
    models/
      schemas.py         # Pydantic models for all three datasets
    db/
      migrations/        # Alembic migrations for PostgreSQL

Example 2: Converting Arena Votes to DPO Training Data

User: "I have 50,000 pairwise preference votes from my LLM arena. How do I convert them into a DPO dataset?"

Approach:

1. Load conversations and votes datasets, join on conversation_id
2. Filter out ties (they carry no preference signal for DPO)
3. For each vote with a clear winner, extract the user prompt, the chosen response (winner), and the rejected response (loser)
4. Detect and tag language per conversation using fasttext
5. Apply the PII filter — exclude any conversation flagged as containing personal data
6. Export in the standard DPO format: {"prompt": ..., "chosen": ..., "rejected": ...}

Output:

{
  "prompt": "Quels sont les avantages et inconvénients du nucléaire en France ?",
  "chosen": "L'énergie nucléaire représente environ 70% de la production électrique française. Parmi les avantages : faibles émissions de CO2, indépendance énergétique, coût de production stable...",
  "rejected": "Le nucléaire c'est bien parce que ça produit beaucoup d'énergie. Les inconvénients c'est les déchets.",
  "language": "fr"
}

Example 3: Adding a New Language to an Existing Arena

User: "Our arena works for French. We want to add Swedish and Lithuanian support as compar:IA did with their European expansion."

Approach:

1. Add language detection at the prompt ingestion step — use fasttext lid.176.bin to auto-tag each conversation
2. Configure inference providers that support the target languages — verify each model's training data coverage
3. Adjust the PII classifier prompt to handle Swedish/Lithuanian PII patterns (personnummer, asmens kodas)
4. Create per-language leaderboard views by partitioning Bradley-Terry fitting by detected language
5. Update the topic categorization model or prompts to handle multilingual input
6. Set up separate dataset export splits per language

Output: The leaderboard now shows separate rankings per language, and the exported HuggingFace datasets include a language column for filtering.

Best Practices

• Do: Randomize left/right model assignment on every comparison to prevent position bias. compar:IA found this is a significant confound if not addressed.
• Do: Use full-conversation exclusion for PII rather than token-level redaction. The 5% data loss is worth the privacy guarantee. Span-level anonymization leaves residual re-identification risk.
• Do: Support multi-turn conversations before voting. Single-turn comparisons miss how models handle context, follow-ups, and clarifications.
• Do: Record both conversation-level votes AND message-level reactions. Conversation votes give DPO pairs; message reactions enable turn-level reward models.
• Avoid: Constraining prompts with pre-set categories or templates. compar:IA found that fewer than 6% of users used suggested prompts — unconstrained input yields more realistic, diverse data.
• Avoid: Using raw win rates for leaderboards. Bradley-Terry properly handles transitive inconsistencies and uneven matchup frequencies. Raw percentages mislead when some models face tougher opponents.
• Avoid: Assuming API providers serve the exact model advertised. Quantization, preprocessing, and routing logic on the provider side can alter outputs. Log provider metadata and flag known discrepancies.

Error Handling

• Inference timeout or failure: If one model fails to respond, do not show the comparison. Return the user to the prompt screen with an apology. Never show a one-sided comparison — it biases against the failed model.
• Inconsistent system prompts: Open-weight models via third-party providers may lack system prompts while proprietary APIs inject them. Standardize a minimal system prompt across all models or document the inconsistency in dataset metadata.
• PII filter false positives: Some legitimate conversations about public figures or fictional characters may trigger the PII classifier. Provide a manual review queue and a public reporting form for users to contest exclusions.
• Language detection errors: Short prompts (under 10 words) cause unreliable language detection. For prompts below a character threshold, either skip language tagging or use a secondary classifier.
• Vote spam or gaming: Monitor for patterns — same IP casting rapid votes, always choosing the same side. Apply rate limiting and flag anomalous voting patterns for manual review.

Limitations

• Self-selection bias: Arena users skew toward digitally literate, tech-savvy populations. The resulting preference data may not represent general population preferences. This cannot be fully corrected post-hoc.
• Verbosity and style bias: Human evaluators systematically prefer longer, more confident, better-formatted responses regardless of factual accuracy. The collected preferences encode these biases, which then propagate into DPO/RLHF-trained models.
• Proprietary model licensing: Responses from proprietary models (GPT-4, Claude, Gemini) cannot be used for training due to terms of service — only for evaluation and analysis. This limits the usable preference pairs for fine-tuning.
• Pairwise evaluation hides absolute quality: A model that wins every comparison might still produce mediocre outputs if all models are weak. Bradley-Terry measures relative strength, not absolute capability.
• Environmental cost estimates are approximate: For proprietary models, architecture and parameter count are estimated from public information. Actual energy consumption may differ significantly.

Reference

Paper: compar:IA: The French Government's LLM arena to collect French-language human prompts and preference data — Termignon et al., 2026. Focus on Section 3 (platform architecture and data schema), Section 4 (Bradley-Terry ranking methodology), and Section 5 (privacy filtering pipeline) for implementation details.