TencentGR-10M: Industrial-Scale GR Dataset

• TencentGR-10M is a large-scale, industrial-grade benchmark dataset designed for generative recommendation by capturing sequential user–item interactions in online advertising.
• It integrates sparse collaborative features with dense multi-modal embeddings from text and image models to comprehensively represent real-world ad interactions.
• The dataset employs a specialized evaluation protocol with weighted metrics and a baseline Transformer model to benchmark performance on high-value actions.

TencentGR-10M is a large-scale, industrial-grade benchmark dataset for all-modality generative recommendation (GR), released as part of the final track of the Tencent Advertising Algorithm Challenge 2025. Built from real, de-identified Tencent Ads logs, TencentGR-10M distinguishes itself through its scale, the integration of rich collaborative categorical identifiers and multi-modal content (such as text and vision embeddings), and a task and evaluation protocol custom-designed for generative modeling of sequential user–item interactions in online advertising (Pan et al., 4 Apr 2026).

1. Dataset Construction and Scale

TencentGR-10M comprises 10,139,575 users, each with up to 100 logged sequential ad interactions, yielding approximately 986 million exposure, click, and conversion events. The ad corpus spans 17,487,676 distinct items, with a global candidate pool of 3,637,720 ads relevant for retrieval ranking. User sequences $S_u$ begin with a user-profile token, followed by item-interaction tokens $x_{u,1}, ..., x_{u,T_u}$ (average $T_u = 97.29$). Each $x_{u,t}$ is annotated with an action-type label $$r_{u,t} \in \{0 = \mathrm{exposure}, 1 = \mathrm{click}, 2 = \mathrm{conversion}\}$$.

The prediction target construction is aligned with industry-standard conversion attribution: if a conversion post-$t_\mathrm{begin}$ can be attributed within the attribution window, this conversion is mapped back to the causal click and impression; otherwise, the earliest post-$t_\mathrm{begin}$ click (and its impression) is taken as the target. Conversions thus appear both within the history and as possible targets. The overall action-type distribution is 94.63% exposures, 2.85% clicks, and 2.52% conversions.

All user and item identifiers, along with sensitive content (names, text, videos, raw images), are de-identified or removed; only hashed categorical IDs and pre-computed embeddings derived from content remain. No user sequence exceeds 100 item tokens.

2. Feature Schema and Modalities

TencentGR-10M merges sparse collaborative (ID-based) features with dense, pre-extracted multi-modal embeddings for each ad, yielding a fully all-modality dataset.

Sparse Features

Ad-side categorical fields (feature IDs 100–122) include single-valued features such as format (6 values, 99.96% coverage), advertiser type (53 values), campaign ID (173,463 values), business line (30 values), device type (33 values), and category (e.g., second-level category, 988 values, 29% coverage). User-side features (IDs 103–110) include age-bucket (87 values), gender (2), region (7), long-term and short-term interest tags (multi-valued, 14 and 19 tags, respectively), device OS (3), and other demographic categorizations.

Multi-Modal Embeddings

Each ad interaction can include up to six embeddings from high-capacity pretrained models:

Embedding ID	Model	Modality	Size	Output Dim
81	Bert-finetune	Text	0.3B	32
82	Conan-embedding-v1	Text	0.3B	1024
83	gte-Qwen2-7B-instruct	Text	7B	3584
84	hunyuan_mm_7B_finetune	Image	7B	32
85	QQMM-embed-v1	Image	8B	3584
86	UniME-LLaVA-OneVision-7B	Image	8B	3584

Text models 81 and 83 are further finetuned with contrastive loss using Tencent’s click/conversion logs. Most embeddings have $>90\%$ coverage in the dataset.

Record Schema

Each interaction record consists of hashed 32-bit categorical IDs, an 8-bit action code, and float32 multi-modal vectors. For each user entry in Parquet format, fields include user_id, user_feats, a sequence of item-interaction dictionaries (with ad_id, campaign_id, etc.), candidate_pool of ads, target_ad, and target_action.

3. Data Format, Access, and Partitioning

TencentGR-10M is distributed via HuggingFace at https://huggingface.co/datasets/TAAC2025/TencentGR-10M as compressed Parquet files. Directory layout follows: train/part-xxxx.parquet, validation/*.parquet, test/*.parquet. Users are shuffled and split chronologically by timestamps $t_\mathrm{begin}$, $t_\mathrm{end}$: 80% for training, 10% for validation, 10% for test. Each Parquet partition consists of approximately 100,000 full user sequences.

The candidate pool and prediction targets are provided explicitly, consistent with the all-modality recommendation protocol. A private test set, withheld by organizers, is used for final evaluation; only training and validation labels are public.

4. Evaluation Protocol

The evaluation framework weights high-value conversions more heavily, employing “weighted” variants of standard metrics. For user $x_{u,1}, ..., x_{u,T_u}$0, ground-truth $x_{u,1}, ..., x_{u,T_u}$1, and predicted $x_{u,1}, ..., x_{u,T_u}$2:

• Action weights: $x_{u,1}, ..., x_{u,T_u}$3, $x_{u,1}, ..., x_{u,T_u}$4, $x_{u,1}, ..., x_{u,T_u}$5.
• Weighted HitRate@K:

$x_{u,1}, ..., x_{u,T_u}$6

• Weighted DCG@K, IDCG@K, and NDCG@K defined by substituting the above weighting into standard definitions, with final leaderboard score:

$x_{u,1}, ..., x_{u,T_u}$7

Weights are calibrated such that both terms have comparable magnitude.

The private test set labels are not disclosed; only train and validation sets feature public labels for model development and hyperparameter tuning.

5. Baseline Model and Training Objective

The baseline for TencentGR-10M is a single-block ($x_{u,1}, ..., x_{u,T_u}$8) causal Transformer with hidden size $x_{u,1}, ..., x_{u,T_u}$9, a single attention head, and dropout of 0.2. Input sequences can be up to 1 user token plus 100 item tokens.

Input Representation

Sparse fields are embedded via learnable lookups; continuous multi-modal embeddings are concatenated. User token $T_u = 97.29$0 is obtained through a multilayer perceptron (MLP) over all user-side embeddings. Item token $T_u = 97.29$1 uses an MLP over both ad-side categorical fields and all available multimodal embeddings. Positional encodings $T_u = 97.29$2 are added, and the resulting sequence $T_u = 97.29$3 is passed through the Transformer layer to yield $T_u = 97.29$4 with user state $T_u = 97.29$5.

Contrastive Loss

Training employs the InfoNCE loss over one positive target and sampled negatives:

$T_u = 97.29$6

where $T_u = 97.29$7, and $T_u = 97.29$8, $T_u = 97.29$9, $x_{u,t}$0 matching the evaluation weighting.

Inference

All item embeddings are precomputed and indexed via ANN (e.g., Faiss). For each user, the most recent state $x_{u,t}$1 is used to fetch the top-K nearest items. Baseline code and instructions are available at https://github.com/TencentAdvertisingAlgorithmCompetition/baseline_2025. No absolute performance numbers for the 10M track baseline are published.

6. Key Findings from the Challenge

Analysis of top-ranked solutions in the final round identified several reproducible trends:

• Action-conditioning: Modeling exposure, click, and conversion as explicit conditioning signals (via gated FiLM layers or special tokens) enables superior intent disentanglement.
• Semantic IDs via quantization: Multi-modal quantization (e.g., RQ-KMeans, SVD-based quantizers) transforming continuous embeddings to discrete codes enhances long-tail item coverage and reduces memory footprint.
• Hierarchical time/session features: Employing representations such as absolute timestamps, inter-event gaps, session segmentation, cross-day indicators, and Fourier decompositions yields measurable improvements.
• Contrastive learning scale: Utilizing large-scale negative banks (order $x_{u,t}$2 negatives per batch) provides greater performance gains than increasing model depth/width under constant compute.
• Two-stage training: Pretraining on exposures for self-supervised next-item prediction, followed by finetuning on clicks/conversions, better aligns models with high-value action distributions.
• Efficiency techniques: Adoption of mixed-precision (AMP), FlashAttention, grouped GEMMs, KV cache optimizations, separate sparse/dense optimizers, and single-model (non-ensembled) submissions endows implementations with latency and memory properties suitable for industrial environments.

TencentGR-10M and its companion codebase aim to provide a reproducible and challenging foundation for research on all-modality generative recommendation in advertising at industrial scale (Pan et al., 4 Apr 2026).