Amazon 2018 raw data no longer downloadable — could you share a mirror and/or your DeepSeek preference pseudo-labels?

[SuFame920]

Hi, thanks for the great work and for open-sourcing MiniOneRec.

I'm reproducing the pipeline and ran into a blocker on the data side.

Problem

The official Amazon Reviews 2018 dataset (UCSD / Julian McAuley's host) appears to be down / no longer
downloadable — the original links time out and I can't retrieve the raw review data.

This matters specifically for the preference / "thinking" alignment task: as I understand the pipeline, the raw user
reviews are required as input to prompt DeepSeek and generate the preference pseudo-labels used for that part
of training. Without the raw data I cannot regenerate these labels, and without the labels I cannot fully reproduce the
paper's reported results.

Requests

1. Raw data mirror — Would you be able to provide a working download link (e.g., a cloud drive / HF dataset mirror)
   for the raw Amazon 2018 data you used, so the preprocessing → pseudo-label pipeline can be run end to end?
2. Pre-generated pseudo-labels — Alternatively (or additionally), could you share the DeepSeek-generated preference
   pseudo-labels from your own run? That would let people reproduce the preference task directly even while the original
   source is unavailable.

Either one would unblock reproduction. Thanks a lot in advance!

[SuFame920]

Quick update — while waiting on the data question, I went ahead and reproduced the core pipeline (SID → SFT → GRPO) on a
single GPU with a Qwen2.5-1.5B backbone, on the Industrial_and_Scientific split. A few takeaways:

• The pipeline reproduces cleanly end to end, and RL (GRPO) consistently improves over SFT across all metrics, with
  the largest gains at the top ranks (e.g. NDCG@5 +4.6%, NDCG@3 +5.8%) — exactly what the rank-aware reward design predicts.
• Even at 1.5B, the reproduction reaches roughly 81–88% of the reported numbers, e.g. HR@5 0.1112 / NDCG@5 0.0889 /
  HR@10 0.1401 / NDCG@10 0.0981, landing it between strong LLM baselines (above BIGRec/LCRec/TIGER) and the full model. The
  remaining gap is well explained by model scale (1.5B vs the larger backbone) and the missing preference/"thinking" task.

Really impressive and well-engineered work — the rank-aware reward + constrained beam search combination is elegant and
clearly does the heavy lifting. Thanks for open-sourcing it!

(My reproduction, with code + results + plots: https://github.com/SuFame920/MiniOneRec-Reproduction)

The original blocker still stands though: I can't run the preference/"thinking" task without the raw reviews to generate
the DeepSeek pseudo-labels, so a data mirror and/or your pre-generated labels would be a huge help whenever convenient.