Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
ReRankID is derived from the gated OpenOneRec/OpenOneRec-RecIF dataset, which contains anonymized real user histories and item captions. By requesting access you agree to attribute OpenOneRec/OpenOneRec-RecIF as the source and not to attempt to re-identify any user.
Log in or Sign Up to review the conditions and access this dataset content.
ReRankID
Reasoning Reranking with Semantic IDs as the Sole Item Representation
Zhenwei Tang1,2, Jintao Liu1, Di Bai1, Xidong Wu1, Xinwu Cheng1
1Google 2University of Toronto
ReRankID is a controlled testbed for reasoning reranking in which a model sees a user history and a candidate slate only as Semantic IDs (SIDs), writes a rationale, and ranks the slate, while item rendering, history length, and slate size are varied one at a time. Existing SID models often copy the input order instead of reranking, and readable item text does not repair the failure; supervised fine-tuning (SFT), rejection-sampling fine-tuning (RFT), on-policy distillation (OPD), and reinforcement learning with verifiable rewards (RLVR) all elicit the missing behavior. The strongest model, the 8B model trained with a graded ranking reward, matches strong text-reading LLMs whose prompts are about 20x longer. This repository holds the benchmark, the data to retrain every post-training method, and the per-user outputs behind every table and figure; prompts and rationales are in Chinese, as in the source data.
Contents
| Paper | Path | What it is |
|---|---|---|
| Sec. 3, App. A | data/benchmark/ |
bench.parquet: 800 users (400 dev, 400 test), each with one positive, a 100-item candidate pool, history, SIDs, and captions. pools.parquet: partition of all 162,074 RecIF users. stats.json: construction statistics |
| Sec. 5, App. C.1 | data/sft/ |
12,000 Gemini-3.6-Flash hindsight annotations and the SID-only LLaMA-Factory set (11,031 train, 960 dev) |
| Sec. 5, App. C.2 | data/rft/onerec-{1.7b,8b}/ |
1,940 and 1,912 verifier-accepted self-generated traces |
| Sec. 5, App. C.3 | data/opd/ |
6,000 label-free verl prompts and 400 dev prompts |
| Sec. 5, App. C.4 | data/rlvr/ |
6,000 verl prompts with reward ground truth and 400 dev prompts, shared by the NDCG@5 and Hit@1 rewards |
| Sec. 4, Table 1, Fig. 2 | results/diagnosis/ |
Random, caption ranker, three text references, both OneRec models under the SID, Text, and Both renderings (--render sid, text, sid_text), the history and slate sweeps, and prompt token counts |
| Sec. 5, Table 2, Fig. 2 | results/posttraining/ |
Each method and scale on the 400 test users (test/), its dev selection checkpoints (dev/), and the history and slate sweeps of the 8B rlvr-ndcg5 model |
| App. D.3 | results/audits/stability/ |
Gemini runs under new display orders, prompt rewordings, and decode repeats, with both audit reports |
| Sec. 5, Fig. 3, App. D.4 | results/audits/rationale-judge/ |
Pairwise rationale verdicts for 100 users and the corruption controls |
results/INDEX.json maps every result file to the tables and figures that use it, and each bundle
has a MANIFEST.json with the size and SHA-256 of every file. In data/sft/annotations.jsonl,
split (train, dev, or rejected) and rejection_reason record the outcome of the quality gates in
code/sft/build_dataset.py and the dev hash in code/sft/recipe.py; the code writes the same
split as train.json and dev.json rather than into this file. results/posttraining/rft/onerec-1.7b/dev/init.json is the selected 1.7B SFT dev file
(sft/onerec-1.7b/dev/epoch3.0.json), the RFT initializer.
Names are lower case, hyphens join words (rlvr-ndcg5, gemini-3.6-flash), underscores separate
fields, and an evaluation cell is k<history>_n<slate> (anchor k64_n30, display seed 0;
audit runs add _s<i> for display orders, _p<i> for prompt wordings, and _repeat<i> for
decode repeats). Models are onerec-1.7b (OpenOneRec/OneRec-1.7B-pro) and onerec-8b
(OpenOneRec/OneRec-8B-Pro), methods are sft, rft, opd, rlvr-ndcg5, and rlvr-hit1, and dev
checkpoints carry the paper's labels (epoch3.0, step100, init).
Usage
git clone https://github.com/lilv98/ReRankID && cd ReRankID
export HF_TOKEN=... # an account that has accepted this dataset's terms
python code/download/release.py --install
This downloads both bundles into downloads/rerankid/, checks every file against its manifest,
and copies data/ into the repository's data/, where the scorers and training recipes read it.
import json, pandas as pd
bench = pd.read_parquet("downloads/rerankid/data/benchmark/bench.parquet")
run = json.load(open("downloads/rerankid/results/posttraining/rlvr-ndcg5/onerec-8b/test/k64_n30.json"))
print(run["metrics"]["ndcg@5"], len(run["results"]))
Source and attribution
This dataset is a modified derivative of
OpenOneRec/OpenOneRec-RecIF at
revision 8f7cf2ee0b949e955a87a708d02024687be232c8 (Apache-2.0). We selected users and
partitioned them into evaluation and training pools, built candidate slates from their exposure
logs plus random negatives, rendered histories and slates as SID and caption prompts, and added
teacher annotations, post-training datasets, and model outputs. The SIDs, captions, and user
histories are unchanged from the source. code/download/recif.py,
code/benchmark/build_pools.py, and code/benchmark/build_bench.py rebuild pools.parquet and
bench.parquet from the source row for row; the Gemini-3.6-Flash teacher annotations cannot be
regenerated exactly.
Captions are machine-generated descriptions of public short videos. They can include on-screen text such as account handles, as well as sensitive or offensive topics, and are unchanged from the source.
data/sft/{annotations.jsonl,train.json,dev.json} contain rationales and rankings written by
Gemini-3.6-Flash. results/diagnosis/{gemini-3.6-flash,claude-sonnet-5,gpt-5.6-terra}/,
results/audits/stability/, and results/audits/rationale-judge/ contain outputs of those models
and of Claude-Opus-5, the rationale judge.
License
Apache License 2.0 (see LICENSE). Use is also subject to the terms of the source dataset, and use
of the model outputs listed above may also be subject to the respective providers' terms.
This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program.
- Downloads last month
- 30