You can deploy this model and start earning money today!

sixteensun/
tencentgr-1m-2025-reproduction

Name: tencentgr-1m-2025-reproduction
Availability: InStock

pytorch

safetensors

recommender-systems

generative-recommendation

sasrec

multimodal

taac2025

TAAC2025/TencentGR-1M

cc-by-nc-4.0

Views

No views yet

TencentGR-1M 2025 Generative Recommendation Reproduction

This repository is an independent, non-official reproduction of the 2025 Tencent Ads Algorithm Competition baseline on TencentGR-1M. It packages a verified SASRec-style multimodal sequential recommender, the reproducible training/evaluation code, and aggregate offline metrics.

Results

The offline evaluation holds out the last click from a seeded 90/10 user split, builds histories from strictly earlier events, and retrieves Top-10 items from the official 660k candidate pool.

Variant	maxlen	MM	HR@10	NDCG@10	Score	Final BCE
MM101	101	on	0.0313478	0.0159694	0.0207367	0.2043
no-MM101	101	off	0.0317533	0.0165208	0.0212429	0.2040
MM50	50	on	0.0320827	0.0160503	0.0210203	0.2061
no-MM50	50	off	0.0337046	0.0172092	0.0223228	0.2055

All four prediction files have identical users, targets, and original history lengths row by row. no-MM50 is the best fixed-seed point estimate: +7.65% versus MM101, +5.08% versus no-MM101, and +6.20% versus MM50. The overall 2x2 interaction interval includes zero, and every configuration has one training seed, so this is not evidence that multimodal data is generally harmful. The largest gain is concentrated in the 81+ history slice (63.47% of users).

Model and training setup

Architecture: SASRec-style Transformer sequence encoder with feature fusion
Hidden size: 32
Transformer blocks / heads: 1 / 1
Maximum sequence length: 101 (baseline) / 50 (recent-window variant)
Batch size: 2,048
Optimizer: Adam, learning rate 1e-3, betas (0.9, 0.98)
Epochs: 3
Seed: 2025
Multimodal feature: field 81, 32 dimensions (disabled in no-MM variants)
Training accelerator used for this reproduction: one NVIDIA H200

Two evaluated checkpoints are public. model.safetensors is the audited MM101 baseline (48 tensors). model_nomm50.safetensors is the best fixed-seed point estimate (46 tensors) and must be loaded with --maxlen 50 --disable_mm_emb. Both were converted from their evaluated PyTorch state dicts and verified tensor-by-tensor for identical keys, shapes, dtypes, and values. The original PyTorch checkpoints remain in the private evidence archive.

Repository contents

model.safetensors: evaluated MM101 multimodal baseline state dict
model_nomm50.safetensors: evaluated no-MM50 best-point-estimate state dict
reproduction_config.json: architecture, training, and evaluation contract
metrics/offline_metrics.json: aggregate metrics and audited sample counts
metrics/offline_metrics_no_mm.json: same-protocol no-MM ablation metrics
metrics/ablation_comparison.json: alignment, deltas, and paired uncertainty
metrics/offline_metrics_mm50.json: maxlen-50 multimodal metrics
metrics/offline_metrics_nomm50.json: maxlen-50 no-MM metrics
metrics/four_way_comparison.json: four-way alignment, slices, paired effects, and interaction
docs/: results, resource budget, experiment ledger, resume material, interview kit, and runbook
src/: training, inference, ANN candidate, and offline evaluation code
scripts/: dataset download, validation, and ID-alignment checks
tests/: regression tests for the reproduction additions

The raw TencentGR-1M dataset is not duplicated here. Download it from the official dataset repository.

Quick start

bash

pip install -r requirements.txt
python scripts/download_tencentgr_1m.py /data/TencentGR-1M
python scripts/validate_tencentgr_1m.py /data/TencentGR-1M
python scripts/audit_id_alignment.py /data/TencentGR-1M

hf download sixteensun/tencentgr-1m-2025-reproduction \
  model_nomm50.safetensors --local-dir ./weights

python src/main.py \
  --data_path /data/TencentGR-1M \
  --device cuda:0 \
  --output_dir ./outputs \
  --maxlen 50 \
  --disable_mm_emb

python src/offline_eval.py \
  --data_path /data/TencentGR-1M \
  --checkpoint ./weights/model_nomm50.safetensors \
  --output_dir ./offline_eval \
  --scratch_dir ./offline_eval_scratch \
  --device cuda:0 \
  --maxlen 50 \
  --disable_mm_emb

The reproduction entry points accept both tensor-only PyTorch state dicts and SafeTensors checkpoints. SafeTensors is used for both published checkpoints. Use model.safetensors without the two no-MM50 flags to reproduce the MM101 baseline instead.

Evaluation scope and limitations

These are reproducible offline metrics, not an official leaderboard score.
The candidate pool excludes 19,518 held-out targets that are absent from the official candidate set; 1,745 validation users have no click target.
The checkpoint includes large user/item embedding tables and is tied to the official TencentGR-1M indexing contract.
This is a baseline reproduction intended for research and portfolio use, not a claim of competition placement.

Attribution and license

The implementation is derived from the official 2025 Tencent Ads baseline, licensed under CC BY-NC 4.0. TencentGR-1M is published by TAAC2025 under CC BY 4.0. This repository therefore uses CC BY-NC 4.0 and is restricted to non-commercial use. See LICENSE and ATTRIBUTION.md.