Add StakeBench v1.1.2: raw

.gitattributes

@@ -79,3 +79,8 @@ model_outputs/Qwen3-8B/llm_audit.jsonl filter=lfs diff=lfs merge=lfs -text
 model_outputs/finma-30b/llm_audit.jsonl filter=lfs diff=lfs merge=lfs -text
 polymarket/comments_clean.csv filter=lfs diff=lfs merge=lfs -text
 polymarket/positions_clean.csv filter=lfs diff=lfs merge=lfs -text
+raw/manifold/comments.csv filter=lfs diff=lfs merge=lfs -text
+raw/manifold/user_positions.csv filter=lfs diff=lfs merge=lfs -text
+raw/polymarket/comments.csv filter=lfs diff=lfs merge=lfs -text
+raw/polymarket/user_positions.csv filter=lfs diff=lfs merge=lfs -text
+raw/polymarket/valid_topics.jsonl filter=lfs diff=lfs merge=lfs -text

raw/README.md

@@ -0,0 +1,81 @@
+# StakeBench raw collection files (added in v1.1)
+
+This folder contains the unmodified outputs of the collection scripts, before the cleaning
+steps that produce `polymarket/comments_clean.csv`, `polymarket/positions_clean.csv`,
+`manifold/comments_clean.csv` and `manifold/positions_clean.csv`. They are released so that
+the market-level cleaning and the benchmark instance filters can be audited and re-run under
+other thresholds. All files are UTF-8 without BOM. Collection scripts and cleaning scripts
+are in the code repository under `collection/`
+(https://github.com/PEIYUNHUA/stakebench_code_NeurIPS26).
+
+## Files
+
+| File | Rows | Size | Produced by | Content |
+|---|---|---|---|---|
+| `polymarket/valid_topics.jsonl` | 22,240 events | 40 MB | `get_polymarket_topics.py` | Every closed Polymarket event with total volume >= $10,000 and event-level comment count >= 100, with its member markets (`market_id`, `slug`, `question`, `condition_id`, `volume`, `clob_token_ids`, `end_date`, `resolution`) and the highest-volume market as `best_market`. `polymarket/topics_summary.csv` in the main folder is the flat summary of this file. |
+| `polymarket/comments.csv` | 514,314 | 357 MB | `get_polymarket_comments.py` | Comment crawl output with the commenter's reconstructed position at comment time. Same columns as `comments_clean.csv` plus `category` (always empty; the Gamma API does not return it). Covers 2,070 markets and 24,909 wallets, May 2024 to Mar 2026. |
+| `polymarket/user_positions.csv` | 907,375 | 303 MB | `get_polymarket_position.py` | Activity records from `data-api.polymarket.com/activity` for the 12,771 proxy wallets that hold a position in `comments_clean.csv`, restricted to the 2,056 condition IDs of the comment markets. Includes `TRADE`, `REDEEM`, `MERGE` and `SPLIT` records; Jan 2024 to Dec 2025 (Unix timestamps). |
+| `manifold/valid_markets.jsonl` | 6,083 markets | 3.6 MB | `get_manifold_topics.py` | Every resolved binary Manifold market with volume >= M$10,000, >= 50 unique bettors and at least one comment. `resolution` is YES (2,380), NO (3,528), CANCEL (137) or MKT (38). `manifold/markets_summary.csv` in the main folder is the flat summary of this file. |
+| `manifold/comments.csv` | 119,805 | 55 MB | `get_manifold_comments.py` | Comment crawl output for the 2,926 YES/NO markets that returned at least 10 comments, with each commenter's position at comment time reconstructed from their bets. Original column names (`user_id`, `user_name`, `position_mana`, `likes`); `reply_count` is empty and is back-calculated during cleaning. 6,205 users, Jan 2022 to Mar 2026. |
+| `manifold/user_positions.csv` | 590,904 | 168 MB | `get_manifold_position.py` | Bet records from `api.manifold.markets/v0/bets` for the 2,799 users that hold a position in `comments_clean.csv`, restricted to the 191 cleaned markets; includes redemption and cancelled bets (`is_redemption`, `is_cancelled`). Dec 2021 to Mar 2026. |
+
+Row counts exclude the header line. Comment texts contain newlines, so `wc -l` does not give
+the row count; use a CSV reader.
+
+## How the cleaned files are derived
+
+| Cleaned file | From | Steps (`collection/clean_*.py`) |
+|---|---|---|
+| `polymarket/comments_clean.csv` (514,282 rows) | `raw/polymarket/comments.csv` | drop the empty `category` column; drop markets with an empty question or fewer than 10 comments; reorder columns. Applied to the released raw file these steps remove no rows (all 514,314 rows pass, which is the count in Table 1 of the paper). The shipped cleaned file, unchanged in content since v1.0, has 32 fewer rows (two comments in each of 16 markets, mostly Series-pooled MLB game markets that carry the same two comment IDs) and was written with a different encoding; see the encoding note below. |
+| `manifold/comments_clean.csv` (46,562 rows, 191 markets) | `raw/manifold/comments.csv` | back-calculate `reply_count` from `parent_comment_id`; rename `user_id -> user_address`, `user_name -> pseudonym`, `position_mana -> position_size`, `likes -> reaction_count`; drop markets with an empty question or fewer than 100 comments. |
+| `polymarket/positions_clean.csv` (862,539 rows) | `raw/polymarket/user_positions.csv` | keep `TRADE` records in the comment markets; compute the running position after each trade (`position_side`, `position_shares`, `position_size`, `yes_cum_shares`, `no_cum_shares`, `potential_arbitrage`, `market_price`, `position_market_value`). |
+| `manifold/positions_clean.csv` (387,262 rows) | `raw/manifold/user_positions.csv` | keep bets with `is_redemption == False` and `is_cancelled == False` in the cleaned markets; compute the same running-position columns. |
+
+The 2,469 Polymarket markets reported as "comment-crawl input" in the paper are the
+`best_market_id` entries of the summary file as it stood when the crawl was run
+(`polymarket/progress.json` in the collection output: 2,070 `done`, 399 `skipped` for fewer than
+10 comments). All 2,469 have an event-level comment count of at least 5,018. The released
+`valid_topics.jsonl` and `topics_summary.csv` are the full 22,240-event list at the documented
+thresholds (volume >= $10,000, comments >= 100). For Manifold, the crawl visited the 5,908 YES/NO
+markets of `valid_markets.jsonl`; 2,926 returned at least 10 comments and 2,982 were skipped.
+
+## Encoding note for `polymarket/comments_clean.csv`
+
+`raw/polymarket/comments.csv` is UTF-8 and keeps all characters. The cleaned Polymarket file
+was originally written in GB18030, and characters outside that code page, mainly emoji, were
+written as `?` (13,433 of its rows differ from the raw text for this reason). The Google Drive
+release (1.1.1) still carries that GB18030 file; in this Hugging Face release (1.1.2) the file,
+like `polymarket/market_attrs.csv`, has been re-encoded to UTF-8 without any other change, so
+the `?` substitutions remain. The evaluation code reads the Drive file with `encoding="latin1"`,
+so the models saw mojibake for non-ASCII punctuation (for example `he¡¯s` for `he's`) in those
+rows; this is the input the reported results were obtained on. For analyses that need the
+original text, join `raw/polymarket/comments.csv` on `market_id`, `comment_id` and `created_at`.
+
+## Column reference for files not described in the main README
+
+`polymarket/user_positions.csv`: `proxy_wallet`, `timestamp` (Unix seconds), `conditionId`,
+`type` (TRADE / REDEEM / MERGE / SPLIT), `side` (BUY / SELL), `outcome`, `outcomeIndex`,
+`size` (shares), `usdcSize` (USD), `price`, `transactionHash`, `title`, `slug`, `eventSlug`.
+
+`manifold/user_positions.csv`: `user_id`, `username`, `bet_id`, `created_at` (ISO 8601),
+`contract_id`, `slug`, `question`, `outcome` (YES / NO), `amount` (Mana, signed), `shares`
+(signed), `prob_before`, `prob_after`, `is_redemption`, `is_cancelled`, `is_filled`.
+
+`polymarket/valid_topics.jsonl` (one event per line): `event_id`, `event_title`, `series_id`,
+`comment_count`, `total_volume`, `market_count`, `markets` (list, see above), `best_market`,
+`category` (keyword-assigned at collection time; the benchmark uses `category_validated` from
+the comment files instead).
+
+`manifold/valid_markets.jsonl` (one market per line): `market_id`, `slug`, `question`,
+`outcome_type`, `category`, `resolution`, `resolution_time`, `close_time`, `created_time`
+(Unix milliseconds), `volume`, `unique_bettors`, `total_liquidity`, `last_comment_time`,
+`creator_id`, `creator_username`, `url`.
+
+## Sensitive information
+
+The raw files contain the same identifier types as the cleaned files and nothing else:
+Polymarket wallet addresses and display names, Manifold user IDs and usernames, market creator
+usernames, and on-chain transaction hashes, all publicly visible on the platforms. The raw
+Manifold comment file covers more markets and users than the cleaned file (2,926 markets,
+6,205 users). The statements in the main README under *Responsible AI Statement* apply
+unchanged; do not attempt to de-anonymise users.

raw/manifold/comments.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddc856fd2a33bad1e480cedd67383f2f66c85e3a767db2f6eb6a24811713dd02
+size 55273796

raw/manifold/user_positions.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c77c350e8f8c6e9fb488cfc130853945a9aae847ab53079c971ac0390547d5b
+size 168253700

raw/polymarket/comments.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8cec64d484779eb2db27fd6d2841d185a39c939d5e0d697192e0d8bf5f631ae9
+size 356611077

raw/polymarket/user_positions.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5553cab83ca32ca0f8d3f96e2398e895ae22aa177c81e78304e60654fbd454bf
+size 302855867

raw/polymarket/valid_topics.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4716c7cac5e4f140361d08c38c883af21cfcad9ab72db57c92d71378b4a2353b
+size 40069017