CoolFace
Datasetpublic

budgiesarecooliguess/rejection-sampling

🧠 Amazon Semantic-ID Recommendation + Reasoning — GPT-5.4 regenerated GPT-5.4 enrichment variant of yufan/recsys-genrec-dataset &nbsp;·&nbsp; Reasoning over Semantic IDs Enhances Generative Recommendation This repo mirrors the format of the original yufan/recsys-genrec-dataset, but the LLM-enriched fields are regenerated with GPT-5.4 (Azure OpenAI). The Video Games domain is fully populated. <cat> below is Video_Games. ✅ Currently available config… See the full description on the dataset page: https://huggingface.co/datasets/budgiesarecooliguess/rejection-sampling.

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes7downloads
Dataset Card

<div align="center">

🧠 Amazon Semantic-ID Recommendation + Reasoning — GPT-5.4 regenerated

GPT-5.4 enrichment variant of [`yufan/recsys-genrec-dataset`](https://huggingface.co/datasets/yufan/recsys-genrec-dataset) &nbsp;·&nbsp; Reasoning over Semantic IDs Enhances Generative Recommendation

</div>

This repo mirrors the format of the original yufan/recsys-genrec-dataset, but the LLM-enriched fields are regenerated with GPT-5.4 (Azure OpenAI). The Video Games domain is fully populated.

<cat> below is `Video_Games`.

✅ Currently available

configsplitssourcenotes
`Video_Games_catalog`trainGPT-5.4 regenerateddetailed_description + sid_interleaved_narrative re-created with GPT-5.4; item_id/sid/sid_tokens/title/description/brand reused unchanged from the original.
`Video_Games_reasoning`trainGPT-5.4 regeneratedreasoning_path + integrated_narrative re-created with GPT-5.4; user_id/history_item_*/item_*/reasoning_question/reasoning_reference reused unchanged from the original (49,133 rows, row-aligned).
`<cat>_seqrec`train / validation / testreused from originalRaw interaction sequences — no LLM content, byte-identical to the original.
`<cat>_rl`train / testreused from originalGRPO RL data (SID-history → SID); no LLM enrichment involved.
`general_reasoning`trainreused from originalGeneral reasoning SFT, shared across categories.

▶️ Phase-1 training ready (Video Games)

All four configs Phase-1 Alignment SFT needs are present for Video Games: Video_Games_seqrec, Video_Games_catalog, Video_Games_reasoning, and general_reasoning. The trainer reads its repo from $SIDR_HF_REPO (default yufan/recsys-genrec-dataset), so just point it here and launch:

bash
export SIDR_HF_REPO=budgiesarecooliguess/recsys-genrec-dataset-gpt5.4
cd SIDReasoner
mkdir -p logs
nohup bash phase1_alignment_sft/sft_Qwen3_enrich.sh Video_Games > logs/phase1_launch.out 2>&1 &

📦 Config schemas (unchanged from the original)

configsplitscontent
`<cat>_seqrec`train / validation / testSequential next-item samples — user_id, history_item_title/id/sid, target item_title/id/sid.
`<cat>_catalog`trainItem catalog — item_id, sid, sid_tokens, title, description, brand, detailed_description (GPT-5.4-enriched), sid_interleaved_narrative (GPT-5.4 item story with its SID woven in).
`<cat>_reasoning`trainOne reasoning narrative per user sequence — original user_id/history_item_*/item_* fields plus GPT-5.4 reasoning_path and integrated_narrative; domain-specific auxiliary fields are preserved when present in the original.
`<cat>_rl`train / testGRPO RL data — pre-packaged prompt, reward_model, extra_info.
`general_reasoning`trainGeneral reasoning SFT (messages/metadata stored as JSON strings).

Quick start

python
from datasets import load_dataset

repo = "budgiesarecooliguess/recsys-genrec-dataset-gpt5.4"
catalog   = load_dataset(repo, "Video_Games_catalog")     # GPT-5.4-enriched item universe + SIDs
seqrec    = load_dataset(repo, "Video_Games_seqrec")      # train / validation / test
reasoning = load_dataset(repo, "Video_Games_reasoning")   # GPT-5.4 reasoning + integrated narrative (Phase-1)
rl        = load_dataset(repo, "Video_Games_rl")          # GRPO prompts + rewards
general   = load_dataset(repo, "general_reasoning")       # general SFT