study_config.yaml

# ── Per-deployment study configuration ────────────────────────────────────────
# Copy this file to each HuggingFace Space and edit as needed.
# Secrets (HF_TOKEN, GH_TOKEN, TINKER_API_KEY) must be set as Space Secrets,
# never stored here.

# "preference" : participants compare Product A vs Product B (7-pt preference scale)
# "likelihood" : participants evaluate a single product (7-pt likelihood-to-buy scale)
study_type: preference

# Categories to include. Each entry needs a name and a count.
#
#   Single category (movies only):
#     categories:
#       - name: movies
#         count: 5
#
#   Two categories (mixed):
#     categories:
#       - name: movies
#         count: 3
#       - name: groceries
#         count: 2
#
# For two-category studies the split (3/2 vs 2/3) is automatically alternated
# across users so the overall pool stays balanced.
# The two counts must sum to pairs_per_user.
categories:
  - name: movies
    count: 2

model_variants:
  - name: base
    model_name: "meta-llama/Llama-3.1-8B-Instruct"
    sampler_path: "tinker://69634b92-5bae-593b-93be-3cbba8541817:train:0/sampler_weights/000200"
    prompt_variant:
      personalization: true
      include_bio: true
    count: 2          # items using this variant for odd-numbered users
# counts swap on alternating users:

pair_selection_seed: 42         # Seed for reproducible 50-item pool selection per category
pairs_per_user: 2               # Total items/pairs shown per participant

# Chat constraints — both set to 3 so each participant has exactly 3 real exchanges.
min_turns: 3                    # Minimum exchanges before "done" button is enabled
max_turns: 3                    # Hard cap; input is disabled after this many exchanges

# Prolific  
prolific_completion_code: "CT9QV8FR"
prolific_study_id: "69fd3a38d9fe8adaa3d7ee7d"

# HuggingFace dataset repo where results (JSON + CSV) are uploaded
output_dataset_repo: "ehejin/user_study-preference-personalized_0505_NP3"