leczhang/recenter-personalization

Re-Centering Humans in LLM Personalization — Data

Data for the paper Re-Centering Humans in LLM Personalization (Lechen Zhang, Jiarui Liu, Tal August). Code: https://github.com/orange0629/recenter-personalization

We frame personalization as a three-stage pipeline and collect human judgments for every stage, grounded in real users from WildChat:

Stage	Task	Human judgments	Folder
1	Attribute extraction — is an attribute extracted from a user's history accurate?	5,949 (1,983 attributes × 3 annotators)	stage1/
2	Relevance matching — should an attribute influence the response to a new prompt?	11,919 (3,973 attribute–prompt pairs × 3)	stage2/
3	Personalized generation — is the personalized response better than a generic one, per attribute?	1,101 (367 item–attribute pairs × 3)	stage3/

Each stage also includes the outputs of the five LLMs studied in the paper (Llama-3.3-70B, Qwen3.5-27B, Gemma-4-31B, GPT-5.4, Claude-Sonnet-4.6).

Quick start

from datasets import load_dataset
items = load_dataset("leczhang/recenter-personalization", "stage2_items", split="train")

To use the analysis code, download the whole repository into the code repo's data/ folder:

hf download leczhang/recenter-personalization --repo-type dataset --local-dir data

Many fields are nested (conversations, per-attribute lists), so reading the JSONL files directly with json / pandas is often the most convenient.

Layout

corpus/                                   # Stage-1 inputs at corpus scale
  wildchat_user_attributes.jsonl          # 14,812 WildChat users: raw + merged extracted attributes
  wildchat_assistant_usage_scores.jsonl   # 16,573 users: assistant-like-usage score + LLM reasoning
  wildchat_sampled_user_ids.json          # diversity-based samples of 100 / 1,000 / 5,000 users
  synthetic/{cupid,prefeval,personalens}_conversations.jsonl    # 100 users each
  synthetic/{cupid,prefeval,personalens}_attributes_raw.jsonl   # extracted attributes
stage1/
  items.jsonl                  # 77 users (47 WildChat + 10 each CUPID/PrefEval/PersonaLens), 1,983 attributes
  annotations/ann{01,02,03}.jsonl
  attr_evidence_cache.jsonl    # evidence (conversation, turns) highlighted for annotators
  verifier_predictions/        # Table 2: LLM verifiers (zero-shot / optimized prompt), test split
stage2/
  items.jsonl                  # 164 (LIMA prompt, user) pairs, 41 prompts × 4 users, 3,973 attributes
  annotations/ann{04,05,06}.jsonl
  llm_judgments/<model>.jsonl  # YES/NO + full reasoning for every pair
  sentence_embedding_scores.json   # all-MiniLM-L6-v2 retrieval baseline scores
  grpo/{train,val,test}.parquet    # verl-format data for GRPO
stage3/
  items.jsonl                  # 582 generic-vs-personalized items; item_index < 80 were annotated by humans
  items_gt.jsonl               # the 80 annotated items with per-attribute mean human ratings (reward-model data)
  annotations/ann{04,05,06}.jsonl
  generations/{no_context,gt_attributes}_<model>.jsonl   # generic / personalized responses, 164 per file
  llm_judgments/<model>.jsonl  # open-weight judges: all 582 items; GPT-5.4 / Claude: the 80 annotated items
  mention_check/gpt-5.4.jsonl  # does the personalized response explicitly mention the attribute?

<model> ∈ llama-3.3-70b, qwen3.5-27b, gemma-4-31b-it, gpt-5.4, claude-sonnet-4-6.

Record formats

Users and attributes. user_id is WildChat's hashed_ip, or the persona id for synthetic users. Each entry of merged_attributes is one deduplicated attribute: {attribute, reason, confidence, original_attributes, original_reasons, original_confidences}. In stage1/items.jsonl each attribute also carries ground_truth (1 = accepted by the annotator majority, 0 = uncertain/rejected) and evidence ({conv_idx, turn_idxs}). conversations is a list of {conversation_hash, model, conversation: [[{role, content}, ...], ...]}. For WildChat we keep only roles and text (no IP-derived location, headers or timestamps).

Stage 1 annotations (one record per annotator × user): index (= item_index) and attr_judgments[i].judgment ∈ {ok = makes sense, unsure = maybe, bad = doesn't make sense}, aligned with merged_attributes.

Stage 2 items (stage2/items.jsonl): item_index, prompt_index, prompt_text, user_id, profile_attributes (the user's attributes, each with ground_truth = human-majority relevance, ties → relevant), and items (attributes proposed by the Llama-3.3-70B coarse filter). Stage 2 annotations: index (= item_index) and relevance_judgments: [{attribute, relevant}]. LLM judgments: sample_index (= item_index) and items: [{attribute, relevant, raw_answer, full_output}].

Stage 3 items: item_index, model_tag (generator), prompt_text, attributes: [{attribute, is_relevant}], response_no_context, response_personalized. Stage 3 annotations / LLM judgments: ratings: {attribute: 1–5} in canonical orientation (1 = strongly prefer the generic response, 5 = strongly prefer the personalized response), plus the display_assignment of the A/B sides that the rater saw.

Annotators

Annotators were recruited on Prolific and paid ~$18/hour. Their IDs are replaced by pseudonyms: ann01–ann03 for Stage 1, and ann04–ann06 for Stages 2 and 3.

Sources and licenses

Our annotations and model outputs are released under CC BY-SA 4.0. Derived content keeps the terms of its source:

• WildChat-1M conversations and derived attributes: ODC-BY. WildChat was PII-redacted by its authors. user_id is WildChat's own hashed IP. Some content may be sensitive.
• CUPID (kixlab/CUPID): CC BY 4.0.
• PrefEval (siyanzhao/prefeval_implicit_persona).
• PersonaLens dialogues: generated with the PersonaLens framework (CC BY-NC 4.0), with Claude 3.5 Haiku on Amazon Bedrock.
• LIMA (GAIR/lima): the 41 study prompts appear in the Stage 2 and 3 items.

Citation

@article{zhang2026recentering,
  title   = {Re-Centering Humans in LLM Personalization},
  author  = {Zhang, Lechen and Liu, Jiarui and August, Tal},
  journal = {arXiv preprint arXiv:2606.06614},
  year    = {2026}
}