elismasilva/dod-agent-traces

DOD Agent Traces

This dataset stores lightweight agent traces generated by DOD - Deploy or Draw, a multiplayer UNO-style game built for the Build Small Hackathon.

The traces are meant to show how the AI parts of the game participate in live gameplay:

• Nemotron bot turns: the AI opponent receives the current board state and chooses whether to play a card or draw.
• IT Director reactions: the LLM generates a short contextual reaction to the card that was just played and the active production crisis.

These records are not training data for a fine-tuned model. They are runtime traces that document how the AI agent is used inside the game loop.

What Is Recorded

Each line in dod_agent_traces.jsonl is a JSON object with this top-level structure:

{
  "id": "unique_trace_id",
  "timestamp": "2026-06-14T12:00:00.000000+00:00",
  "event": "nemotron_turn",
  "model": "NVIDIA-Nemotron-3-Nano-4B-GGUF",
  "status": "llm_play_applied",
  "input": {},
  "output": {},
  "metadata": {}
}

If an error occurs and the game has to use a fallback path, the trace can also include:

{
  "error": "short error description"
}

Event Types

nemotron_turn

Recorded when Nemotron is asked to make a gameplay decision.

The input includes:

• the active card stack and category;
• current match metrics, such as crisis resolution and director panic;
• Nemotron's hand;
• which cards are legally playable according to the backend rules;
• whether at least one playable card exists.

The output can include:

• the raw LLM response;
• the parsed decision;
• whether a fallback was used.

Possible statuses include:

• llm_play_applied: the model selected a valid playable card.
• llm_draw_applied: the model chose to draw when no legal card was available.
• llm_invalid_card_fallback: the model selected an invalid card and the backend used local fallback rules.
• llm_draw_rejected_fallback: the model chose to draw despite having playable cards, so the backend rejected the move and used local fallback rules.
• remote_error_fallback: the remote LLM call failed and the backend used local fallback rules.
• fallback_failed: both the LLM call and fallback handling failed.

director_reaction

Recorded when the game asks the LLM to generate the IT Director's reaction.

The input includes:

• the played card;
• the card type;
• card effect details;
• recent Director quotes;
• the current production crisis.

The output can include:

• the raw LLM response;
• the generated quote in English and Portuguese;
• whether a fallback quote was used.

Possible statuses include:

• quote_generated: the LLM generated and validated the Director quote.
• fallback_quote_applied: the LLM failed or returned an invalid quote, so the game used a fixed crisis fallback quote.
• fallback_quote_failed: both quote generation and fallback handling failed.

Privacy And Scope

The trace payload is intentionally compact. It focuses on gameplay and model behavior, not personal data.

It does not intentionally store:

• player emails;
• Hugging Face access tokens;
• API keys;
• browser cookies;
• raw audio files.

Some traces may include player-facing game text, card names, crisis names, model output, and fallback/error summaries. The purpose is to make the agent behavior inspectable for the hackathon and for later technical writeups.

Source Project

• Live demo: https://huggingface.co/spaces/build-small-hackathon/dod-uno
• Development repository: https://github.com/DEVAIEXP/doduno
• Blog post: https://huggingface.co/blog/elismasilva/dod-uno-en
• Demo video: https://youtu.be/QahSgSKtvNQ