---
language:
- fr
task_categories:
- question-answering
- text-generation
pretty_name: "delhIA — Conversation Runs"
tags:
- dialogue-evaluation
- tool-use
- french
- weather
- benchmark
- agent-evaluation
- openai-format
license: cc-by-4.0
---

# delhIA — Conversation Runs

This dataset contains **150 conversation runs** on the delhIA Météo-France domain:
50 evaluation scenarios × 3 LLM assistants, produced by a fully automated multi-turn simulation framework.

Conversations are stored in **OpenAI Chat Completions format** (roles: `assistant`, `user`, `tool`).

## Dataset at a Glance

| | |
|---|---|
| **Runs** | 150 (50 tasks × 3 models) |
| **Models** | `gpt-4o`, `qwen3-32b`, `qwen3-8b` |
| **Domain** | Météo-France weather assistant |
| **Language** | French |
| **Format** | JSON — OpenAI messages format |

## Models

| Model | Provider |
|---|---|
| `gpt-4o` | OpenAI |
| `qwen3-32b` | Alibaba |
| `qwen3-8b` | Alibaba |

All models used the same assistant system prompt and the same 50 scenarios.

## Data Structure

Each record in `data/runs.json`:

```json
{
  "task_id": "0_baseline",
  "model": "gpt-4o",
  "status": "FINISHED: ...",
  "telemetry": {
    "token_usage": {"prompt_tokens": 6348, "completion_tokens": 233, "total_tokens": 6581},
    "total_calls": 6,
    "total_tool_calls": 2,
    "average_latency": 2.99
  },
  "messages": [
    {"role": "assistant", "content": "Bonjour ! Comment puis-je vous aider ?"},
    {"role": "user", "content": "Bonjour, je voudrais vérifier les alertes..."},
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [{
        "id": "call_xxx", "type": "function",
        "function": {"name": "meteo_alerts", "arguments": "{\"domain\": \"75\"}"}
      }]
    },
    {"role": "tool", "tool_call_id": "call_xxx", "content": "Alertes de vigilance..."},
    {"role": "assistant", "content": "Pour le département 75, toutes les alertes sont vertes."},
    {
      "role": "user",
      "tool_calls": [{
        "id": "call_yyy", "type": "function",
        "function": {"name": "end_simulation", "arguments": "{\"reason\": \"...\"}"}
      }]
    }
  ]
}
```

**Notes:**
- `end_simulation` is a simulation control tool called by the user simulator to signal task completion — not a Météo-France tool.
- `tool` role messages contain raw Météo-France API responses.

## Usage

```python
import json

with open("data/runs.json") as f:
    runs = json.load(f)

# Filter by model
gpt4o_runs = [r for r in runs if r["model"] == "gpt-4o"]

for run in gpt4o_runs[:1]:
    for msg in run["messages"]:
        role    = msg["role"]
        content = msg.get("content") or "[tool call]"
        print(f"{role}: {content[:80]}")
```

## Companion Dataset

The 50 evaluation scenarios (task definitions, expected tool calls, user instructions) are at
**[Jourdain/delhIA-scenarios](https://huggingface.co/datasets/Jourdain/delhIA-scenarios)**.

## License

[Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/)