manifold/README.md

Manifold UserModeling Benchmark (Belief WIP)

This repository mirrors the Polymarket benchmark layout and currently focuses on Belief-layer data preparation.

Folder Layout

• dataset/Belief/source/
• dataset/Belief/splits/
• dataset/Belief/support_set/
• benchmark/Belief/{direct,profile,retrieval}/ (placeholders for next stage)
• metadata/

---

1) Source Tables (dataset/Belief/source/)

data.csv

User-market final belief master table (one row per userId*marketId).

Key columns:

• userId, marketId
• options_n
• dominant_option, dominant_net, dominant_share
• candidate_options_json
• candidate_net_json
• candidate_bets_json

Use this table as the primary supervised target source for Belief tasks.

option_accumulate.csv

Candidate-level accumulate table (one row per userId*marketId*option).

Key columns:

• option_key (ANS:<answerId> for multi-choice, SIDE:YES/NO for binary-like)
• accumulate_net_amount
• bets_count

Use this when you need per-candidate decomposition rather than only the dominant option.

---

2) Query Splits (dataset/Belief/splits/)

train.csv, validation.csv, test.csv

Rows are query userId*marketId samples split by per-user market last-bet time chronology.

split_summary.txt

Split counts summary.

wallet_ids_experiment_2000.csv

The sampled query user pool (2000 users, bucketed by markets_per_user).

---

3) Support Set (dataset/Belief/support_set/)

support_users.csv

Support user list and user-level stats.

support_belief_final.csv

Support users' final user*market belief summary (dominant/candidate aggregate view).

support_belief_option_accumulate.csv

Support users' candidate-level accumulate records.

support_user_bets_chrono.csv

Support users' deduplicated bet events, sorted by:

1. userId
2. createdTime
3. betId

Key columns include:

• userId, createdTime, betId, marketId
• outcome, answerId, amount, shares
• probBefore, probAfter, isApi

support_user_bets_index.csv

Per-user row-range index into support_user_bets_chrono.csv for fast retrieval.

Columns:

• userId
• start_row_in_events_csv_1based_excluding_header
• end_row_in_events_csv_1based_excluding_header
• num_bets

---

How to Use (Current Stage)

Belief supervised training

• Use dataset/Belief/splits/{train,validation,test}.csv as query samples.
• Use dominant_option (or candidate distribution fields) from source as labels.

Retrieval-style setup

• Query side: split files (train/val/test).
• Support side:
  • User-level belief memory: support_belief_final.csv
  • Event-level memory: support_user_bets_chrono.csv + support_user_bets_index.csv

---

Notes

• Current files are CSV for transparency and easier inspection.
• Next stage will generate benchmark/Belief/direct, profile, and retrieval artifacts aligned with Polymarket format.