traceverse-community/dineout-evals

Dineout — SmolTrace Tasks

Part of the TraceVerse Community open evaluation & observability stack.

A SmolTrace-format evaluation suite of 100 tasks for agents calling dine-in / reservation MCP servers. Pairs with the dineout-mcp Space.

---

Dataset info

• Tasks: 100
• Format: SmolTrace evaluation format
• Tools covered: search_venues, view_venue_details, check_availability, book_table, cancel_booking
• Companion MCP server: traceverse-community/dineout-mcp
• License: Apache 2.0

Usage with SmolTrace

from datasets import load_dataset

dataset = load_dataset("traceverse-community/dineout-evals")
# smoltrace-eval --model openai/gpt-4o-mini --dataset-name traceverse-community/dineout-evals

Schema

Column	Type	Description
id	string	dineout_<tool>[_<suffix>]_batch<n>_<m>
prompt	string	Task description
expected_tool	string	One of the 5 dineout tools
difficulty	string	easy / medium / hard
agent_type	string	tool or code
expected_keywords	list[string]	2-7 keywords expected in plan/output

Example rows

# Easy
{
    "id": "dineout_search_venues_batch0_0",
    "prompt": "Find Continental restaurants in Bangalore I can book a table at.",
    "expected_tool": "search_venues",
    "difficulty": "easy",
    "agent_type": "tool",
    "expected_keywords": ["Bangalore", "Continental", "search", "venues"]
}

# Hard / multi-step
{
    "id": "dineout_book_table_search_venues_check_availability_batch12_304",
    "prompt": "Find a top-rated North Indian place in Delhi and book a table for 4 this Saturday at 8 PM.",
    "expected_tool": "book_table",
    "difficulty": "hard",
    "agent_type": "code",
    "expected_keywords": ["Delhi", "North Indian", "4", "Saturday", "8 PM"]
}

---

Prompt template

Save as prompt_template.yaml and pass via system_prompt_path to smolagents.ToolCallingAgent or CodeAgent.

system_prompt: |-
  You are an expert restaurant-reservations assistant. You have MCP tools to search venues, view venue details, check availability, book tables, and cancel bookings.

  Booking lifecycle: search_venues → check_availability (returns slot_ids) → book_table (with that slot_id) → optionally cancel_booking. Pass the slot_id verbatim from check_availability — do not reconstruct.

  Each step is Action (tool call) → Observation. Finish with `final_answer`.

  Example:

  Task: "Book a table for 2 at a Mediterranean place in Mumbai for Saturday 9 PM."
  Action:
  {"name": "search_venues", "arguments": {"city": "Mumbai", "cuisine": "Mediterranean"}}
  Observation: {"results": [{"id": "v-303", "name": "Bandra Olive Bar", "rating": 4.5}, ...]}
  Action:
  {"name": "check_availability", "arguments": {"venue_id": "v-303", "date": "2026-05-09", "party_size": 2, "time_preference": "dinner"}}
  Observation: {"slots": [{"slot_id": "s-v-303-2026-05-09-2100", "time": "21:00", "available": true}, ...]}
  Action:
  {"name": "book_table", "arguments": {"venue_id": "v-303", "slot_id": "s-v-303-2026-05-09-2100", "party_size": 2, "guest_name": "Guest"}}
  Observation: {"booking_id": "b-12345", "status": "confirmed"}
  Action:
  {"name": "final_answer", "arguments": {"answer": "Booked at Bandra Olive Bar, 9 PM Saturday, party of 2. Booking ID b-12345."}}

  Available tools:
  {%- for tool in tools.values() %}
  - {{ tool.to_tool_calling_prompt() }}
  {%- endfor %}

  Rules: always tool-call; pass values not variables; never pass real PII; finish with final_answer.

planning:
  initial_plan: |-
    Survey known facts, list facts to look up, then write a step-by-step plan ending with <end_plan>.
    Task: {{task}}

system_prompt_code: |-
  You are an expert restaurant-reservations assistant solving tasks with Python code. Use Thought / Code / Observation cycles, finish with `final_answer(...)`.

  Tools (callable as Python functions):
  {{code_block_opening_tag}}
  {%- for tool in tools.values() %}
  {{ tool.to_code_prompt() }}
  {% endfor %}
  {{code_block_closing_tag}}

  Rules: Thought + Code each step; pass slot_id verbatim from check_availability; never pass real PII; allowed imports: {{authorized_imports}}.

---

Reproducing this dataset

hf clone datasets/traceverse-community/dineout-evals
cd dineout-evals
python3 generate.py > tasks.jsonl

---

Part of the TraceVerse Community stack

Project	Where
Reference agents + docs	food-delivery-agents
MCP server (this dataset's companion)	dineout-mcp
Trace viewer	TraceMind
Instrumentation	genai-otel-instrument
Evaluation framework	SmolTrace

Sibling datasets:

• food-delivery-evals
• instamart-evals

Citation

@misc{traceverse-dineout-evals,
  author    = {TraceVerse Community},
  title     = {Dineout Evals: a SmolTrace evaluation dataset for dine-in / reservation MCP agents},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/traceverse-community/dineout-evals}
}

Acknowledgement

SmolTrace format conventions originated in MCP-1st-Birthday's synthetic data work via the TraceMind MCP Server.