sonrayll/social-routing-bench

Social Routing Bench — subreddit_dynamic

Prebuilt data for Social Routing Bench: given a Reddit post title, pick the subreddit it was posted to from a runtime-defined candidate set that changes for every example (4–64 candidates, random / semantic / hard look-alike negatives, candidates shown by name or by anonymous description).

This repository is the exact dataset build the published results use, so you can evaluate a model without rebuilding it.

Use

git clone https://github.com/s0NRAYY/RedditClassificationBench && cd RedditClassificationBench
python -m pip install -e .
hf download sonrayll/social-routing-bench --repo-type dataset --local-dir data/subreddit_dynamic
srb run MODEL --dataset data/subreddit_dynamic --task-config tasks/subreddit_dynamic/medium.yaml --output results/MODEL

Pin a revision (--revision <commit>) for reproducible results.

Files

File	Rows	Contents
posts.parquet	874,928	post_id, subreddit, title, selftext (empty: title-only build), created_utc, split (train / dev / test)
communities.parquet	1,242	community_id, name, public_description, split_group (seen / unseen)
manifests/negative_pools.parquet	3,726	One frozen ordered negative pool per (track, target, difficulty); mining_strategy records how it was built
manifests/instances.parquet	949,206	Evaluation instances: post × difficulty with pool id and display-order seed
build.json	—	Build provenance and counts

How it was built

• Source: titles and community descriptions from TheShadow29/subreddit-classification-dataset (MIT). The public archive has titles only, so this is a title-only build; title+selftext evaluation is rejected by the harness.
• Cleaning: exact normalized-title duplicates removed globally (9,077), which also removes exact crossposts.
• Splits: 20% of communities are unseen (all their posts are test-only). seen communities have post-level train / dev / test splits.
• Negatives: random = seeded permutation of other communities in the track; semantic = seeded permutation of the 128 nearest; hard = exact cosine-similarity ranking. Embeddings: sentence-transformers/all-MiniLM-L6-v2@1110a243. Seen communities use means of seen-train post embeddings; unseen communities use public descriptions only.
• Determinism: benchmark_seed = 20260926. For K candidates the harness takes the target plus the first K−1 negatives, so candidate sets are nested across K; display order derives from (seed, post_id, difficulty, K).

Limitations

• Titles only; many titles fit several communities (about 5% of hard examples, 10% at K = 64, are not answerable from the title alone — LLM-annotated estimate in the repository).
• The source data is old and public, so models may have seen these titles during pretraining.
• A classifier with MiniLM embeddings is disadvantaged on hard negatives, which were mined in the same embedding space.

License and content

The benchmark code and this build are released under MIT, following the upstream dataset. Reddit post titles and community descriptions remain subject to Reddit's terms and the rights of their authors. Do not use this data to identify or profile individual users.