lspcloud/prolific-preferences-dynamic

1# ── Per-deployment study configuration ────────────────────────────────────────2# Copy this file to each HuggingFace Space and edit as needed.3# Secrets (HF_TOKEN, GH_TOKEN, TINKER_API_KEY) must be set as Space Secrets,4# never stored here.5 6# "preference" : participants compare Product A vs Product B (7-pt preference scale)7# "likelihood" : participants evaluate a single product (7-pt likelihood-to-buy scale)8# "model_comparison" : one pair; same participant chats with multiple seller models9#   (order randomized). Use pairs_per_user: 1 and comparison_models: (name, model_name,10#   sampler_path, use_demographics, use_background, personalization).11#   JR1-style checkpoints: uniform_initial_survey: true (preference_initial_uniform).12#   JR2-style (persona buyer wording / simple seller): omit or false (preference_initial).13#   Omit model_variants.14# study_type: preference15 16# Categories to include. Each entry needs a name and a count.17#18#   Single category (movies only):19#     categories:20#       - name: movies21#         count: 522#23#   Two categories (mixed):24#     categories:25#       - name: movies26#         count: 327#       - name: groceries28#         count: 229#30# For two-category studies the split (3/2 vs 2/3) is automatically alternated31# across users so the overall pool stays balanced.32# The two counts must sum to pairs_per_user.33# categories:34#   - name: movies35#     count: 236 37# model_variants:38#   - name: base39#     model_name: "meta-llama/Llama-3.1-8B-Instruct"40#     sampler_path: ""41#     prompt_variant:42#       personalization: true43#       include_bio: true44#     count: 2          # items using this variant for odd-numbered users45# counts swap on alternating users:46 47pair_selection_seed: 42         # Seed for reproducible 50-item pool selection per category48# pairs_per_user: 2               # Total items/pairs shown per participant49 50# Chat constraints — both set to 3 so each participant has exactly 3 real exchanges.51# min_turns: 3                    # Minimum exchanges before "done" button is enabled52# max_turns: 3                    # Hard cap; input is disabled after this many exchanges53 54# Prolific  55prolific_completion_code: "CIE6CQV7"56prolific_study_id: "6a07a5ffe759e03e67f9487c"57 58# HuggingFace dataset repo where results (JSON + CSV) are uploaded59output_dataset_repo: "ehejin/user_study-preference-personalized_0514_comparison_JR1_2" 60 61# ── Example: model_comparison (uncomment and set study_type; comment out model_variants) ──62#63print_model_input: true64study_type: model_comparison65categories:66  - name: movies67    count: 168pairs_per_user: 169min_turns: 370max_turns: 371sampling_temperature: 1.0   # Tinker SamplingParams; seller and all call_model calls72comparison_models:73  - name: base_anonymous74    model_name: "meta-llama/Llama-3.1-8B-Instruct"75    sampler_path: ""76    use_demographics: false77    use_background: false78    personalization: false79  - name: finetuned_JR180    model_name: "meta-llama/Llama-3.1-8B-Instruct"81    sampler_path: "tinker://2fdbf0af-7a75-55a2-aadd-9c6cdf4229d5:train:0/sampler_weights/000060"82    use_demographics: false83    use_background: false84    personalization: false85    uniform_initial_survey: true86  - name: finetuned_JR287    model_name: "meta-llama/Llama-3.1-8B-Instruct"88    sampler_path: "tinker://5e6db03e-85d5-5d3c-95db-8c68e7718be1:train:0/sampler_weights/000120"89    use_demographics: false90    use_background: false91    personalization: false92    uniform_initial_survey: false