RISE retweet baseline / 历史转发预测实验
A tiny CPU model of additional retweets after age 15m or 60m through age 24h in the author-released 2011 SEISMIC corpus. This is a retrospective, selected-cascade research benchmark. It is not a current X views, impressions, verified-Home, follower-growth, text, media or human/AI-origin predictor.
两个截止时点模型，各241参数，共482参数。只用截止前转发计数、首末转发时间与冷帖标记；不用正文、作者身份或未来数据。推理只需Python标准库，训练使用单线程PyTorch CPU。
Real evaluation / 真实结果
165,929 canonical, non-repeated original-ID cascades produce 331,858 paired rows. Two rows from one original are one cascade, not two independent posts. Per cutoff: 59,565 outer training, 12,193 label-overlap purge, 94,171 final evaluation. Inner training/validation/purge: 23,779 / 23,822 / 11,964. Same cascade cutoffs always share the split. Normalization and baselines fit training only; epoch selection uses temporal validation only; both models and baselines freeze before one final evaluation transaction.
| Cutoff → 24h additional RT | Strong log-count bucket median MAE | Tiny MLP MAE | Descriptive reduction |
|---|---|---|---|
| 15m | 63.1849 | 60.3988 | 4.41% |
| 60m | 47.0784 | 45.7117 | 2.90% |
Lower is better; errors are additional-retweet counts. Full five-method, cold/warm, MAE/log-MAE/median/p90/WAPE/RMSE results and training history are in evaluation.json. 15m early-cold MAE is slightly worse (61.9882 vs 61.7230); 60m p90 error is slightly worse (105.1834 vs 105.0000). These are descriptive results on one chronological holdout, not a statistical stability or growth claim. Do not compare them with the separate old two-post views-curve errors 28.3/9.4.
Run inference / 运行推理
hf download LIghtJUNction/RISE-retweet-baseline inference.py model-state.json --local-dir rise-retweet
cd rise-retweet
python inference.py --state model-state.json --cutoff 900 --observed 10 --first 30 --last 800
Output: estimated additional_retweets_by_24h, total_retweets_by_24h, cutoff and explicit historical-corpus scope. first and last are seconds since publication, observed no later than the cutoff; with zero observed retweets, omit both. The command is a demonstration, not a measured forecast of a current account. Predictions are nonnegative point estimates, without calibrated intervals or reliability guarantees.
Reproduce training / 复现训练
The author-hosted dataset is fetched directly from the official source, checked against pinned sizes and SHA-256 values, and kept locally. The preparation and training commands require fresh output directories and refuse to overwrite runs. They do not create a public dataset mirror. Install PyTorch separately from its official distribution; this run used Python 3.14.7 and PyTorch 2.14.0 on CPU.
hf download LIghtJUNction/RISE-retweet-baseline prepare.py train.py --local-dir rise-retweet
cd rise-retweet
python prepare.py --output local-seismic
python train.py plan --input local-seismic/retweet-features.csv --provenance local-seismic/prepared-summary.json --output local-benchmark --expected-input-sha256 f17d46503f68474ef777d59e4398e6c9faba9eb5ee9ed67238b861ae3768763b
python train.py train --plan local-benchmark/plan.json
plan only verifies data and freezes the experiment. train fits CPU models and performs one final evaluation after freezing weights and baselines. Raw rows and predictions remain local. train.py preserves the exact training-source fingerprint, including its original local-storage policy; the separately shared model-state.json has an explicit release-provenance note. Learned weights and normalizers are unchanged. Matching data does not promise bit-identical training across different PyTorch versions or hardware. An independent source-reuse preparation check is recorded in the RISE repository.
Source and limits / 来源和局限
- SEISMIC authors and paper §5.1: 2011 English no-hashtag posts selected using future retweet activity. This excludes the normal population of unsuccessful posts. Early-cold groups in the actual release all have retweets by 24h; ordinary zero-final-traffic performance is unknown.
- 139 scientific-notation/precision-uncertain source-ID entries and all 8 segments of 4 repeated canonical IDs are quarantined. Source exceptions include 1,280 segments with 49 non-original retweets; maximum event time is 604,799 seconds. Downloaded files do not verify the paper's stated 14-day completeness. Labels refer to events recorded in the release through age 24h.
- Author identity and original text/media are unavailable. Author-disjoint and same-topic/content deduplication are unverified. Time separation alone does not prove topic independence or current Chinese-X transfer.
- No explicit dataset license was located. Raw CSV, source IDs, per-cascade features and predictions are not mirrored here. The MIT declaration covers our code and numerical parameter release; it does not relicense the source data.
- Original frozen numerical state fingerprint and code/input fingerprints are recorded in the shared files. Release metadata changes do not change learned weights or normalizers. Standard-library versus Torch float32 inference is tolerance-matched, not bit-identical.
RISE source and measured growth experiments · separate old views-curve prototype