Datasetpublic

MiroShark/social-prediction-market-sim

MiroShark Social + Prediction Market Simulation Agent decisions from MiroShark simulations (GitHub). In each simulation, LLM agents with distinct personas (companies, founders, communities, regulators, commentators) share a Twitter/Reddit-style feed and a Polymarket-style prediction market. Every round, each agent reads the feed (or its portfolio and the open markets) and decides what to do: post, comment, quote, like, follow, buy or sell shares, or do nothing. Each row is one… See the full description on the dataset page: https://huggingface.co/datasets/MiroShark/social-prediction-market-sim.

sourceHugging Facemitupdated 20h agoView on Hugging Face

1likes29downloads

Dataset Card

MiroShark Social + Prediction Market Simulation

Agent decisions from MiroShark simulations (GitHub). In each simulation, LLM agents with distinct personas (companies, founders, communities, regulators, commentators) share a Twitter/Reddit-style feed and a Polymarket-style prediction market. Every round, each agent reads the feed (or its portfolio and the open markets) and decides what to do: post, comment, quote, like, follow, buy or sell shares, or do nothing.

Each row is one agent's decision in one round: the full input it saw, its written reasoning, and the actions it took.

—16 simulations, up to 10 rounds each, about 26 agents per simulation
—8,201 decisions in total (6,563 in the cleaned sft config)
—Social actions and trading actions in one dataset
—Generated by deepseek/deepseek-v4-flash, deepseek/deepseek-v4.1-flash and z-ai/glm-4.7-flash via OpenRouter

As far as we know, this is the first public dataset where the same LLM agents both act on a social feed and trade in a prediction market inside one simulation, released in a format ready for fine-tuning.

Rows per model

Model	`raw`	`sft`	`sft` social	`sft` trade
`deepseek/deepseek-v4-flash:nitro`	5,567	4,445	3,331	1,114
`z-ai/glm-4.7-flash:nitro`	1,253	996	634	362
`deepseek/deepseek-v4.1-flash:nitro`	510	401	203	198
`deepseek/deepseek-v4.1-flash`	446	380	213	167
`z-ai/glm-4.7-flash`	425	341	206	135
Total	8,201	6,563	4,587	1,976

Configs

`sft` (default)

Ready for supervised fine-tuning. One row = one prompt and one target completion.

Column	Description
`row_id`	Index into the `raw` config
`sim_id`, `round`, `agent_id`, `model`	Where the decision came from
`kind`	`social` (feed actions) or `trade` (market actions)
`context`	System prompt: agent persona, rules and allowed actions
`prompt`	What the agent sees this round: the social feed, or its portfolio, the active markets and a sentiment summary
`completion`	The agent's reasoning, ending with a line `ACTION: [...]`
`action_only`	The JSON action list alone

Splits are by simulation, not by row, so no simulation appears in both train (6,133 rows, 13 sims) and test (430 rows, 3 sims).

Rows are removed from sft when:

—the prompt is empty (397 rows), or
—the agent acted on a post, comment or user ID that does not appear in its prompt (1,241 rows).

`raw`

All 8,201 decisions with the structured fields from the simulation logs. Use it to build your own prompt format or to study agent behavior across rounds.

Column	Description
`row_id`, `sim_id`, `run_id`, `round`, `agent_id`, `model`	Identifiers
`persona`	Agent persona and instructions
`allowed_actions`	Actions the agent may call
`portfolio`, `market_state`	Cash, positions and open markets (trading rounds only)
`social_feed`	The feed the agent saw
`reasoning`	The agent's written reasoning
`actions`	JSON string of the actions taken
`result`	Tool result, when logged (rarely present)
`kind`	`social` or `trade`
`empty_prompt`, `invented_id`	Quality flags; rows with either flag set are excluded from `sft`

Action format

json

[{"action_type": "quote_post", "action_args": {"post_id": 54, "quote_content": "..."}},
 {"action_type": "like_post", "action_args": {"post_id": 54}}]

Social actions: create_post, create_comment, quote_post, repost, like_post, dislike_post, like_comment, dislike_comment, follow, trend, refresh, do_nothing. Market actions: buy_shares, sell_shares (with market_id, outcome, amount_usd or num_shares).

Loading

python

from datasets import load_dataset

sft = load_dataset("MiroShark/social-prediction-market-sim")          # sft config
raw = load_dataset("MiroShark/social-prediction-market-sim", "raw")

How this compares

Dataset	Social feed actions	Market trades	Persona agents	Prompt + reasoning + tool calls	SFT-ready
This dataset	Yes (post, comment, quote, repost, like, follow)	Yes (prediction market)	Yes	Yes	Yes
divdata	Posts and comments only	No	Yes	Partial, no prompts kept	No
socsim26-sharedtask	Partial	No (game payoffs)	Yes	Yes, in raw tarballs	No
agent-town-economy	Chat only	Wages and prices	Yes	Partial	No
llm-forecast-bench	No	Yes (prediction market)	Yes	No	No
ForecastBench	No	Forecasts, single agent	No	Rationales	No

Research systems that mix a social feed with trading, such as TwinMarket and StockAgent, release code but not their agent decision logs. The social layer follows the OASIS action set.

This dataset is smaller than some of the social-simulation logs above. Its focus is on complete, trainable rows rather than scale.

Reproducing

scripts/build.py rebuilds both configs from the MiroShark simulation exports: it flags empty prompts and invented IDs, drops them from sft, and splits sft by simulation.

Notes and limitations

—All content is synthetic. Posts, trades and market prices come from simulated agents, not real users or real markets.
—Personas and feeds name real organizations and public X handles as simulation characters. Their statements are model-generated and do not come from those organizations or people.
—Completions are unedited model outputs. They can contain factual errors and overconfident trading claims. Nothing here is financial advice.

License

MIT. Generated with MiroShark. The software license does not apply to generated outputs; this dataset is released under MIT.