debug logs

src/data.py

@@ -177,6 +177,30 @@ def _assign_from_category(category: str, n: int, cfg: dict) -> list:
 
     return assigned
 
+def _assign_variants(cfg: dict, n: int) -> list:
+    """
+    Return a list of n variant dicts (one per item), alternating the
+    personalized/base split across users.
+    """
+    variants = cfg.get("model_variants")
+    if not variants:
+        # Fallback: single variant from old-style config
+        return [{"model_name": cfg["model_name"], "prompt_variant": cfg["prompt_variant"]}] * n
+
+    lock = FileLock(str(_data_dir(cfg) / "variant_counter.lock"))
+    with lock:
+        ctr = _read_counter(_data_dir(cfg) / "variant_counter.txt")
+        _write_counter(_data_dir(cfg) / "variant_counter.txt", ctr + 1)
+
+    # Swap counts on every other user
+    v0, v1 = variants[0], variants[1]
+    if ctr % 2 == 1:
+        v0, v1 = v1, v0
+
+    assigned = [v0] * v0["count"] + [v1] * v1["count"]
+    random.shuffle(assigned)  # interleave so variant order isn't predictable
+    print(f"[VARIANTS] user {ctr}: {[v['name'] for v in assigned]}")
+    return assigned
 
 def _compute_counts(cfg: dict) -> dict:
     """
@@ -268,8 +292,15 @@ def _make_item_slot(item: dict, study_type: str) -> dict:
 def init_state(cfg: dict) -> dict:
     """Build the initial session-state dict for a new participant."""
     n     = cfg["pairs_per_user"]
+    variants = _assign_variants(cfg, n)
     items = assign_items(cfg)[:n]
-
+    for item_slot, variant in zip(items, variants):
+        item_slot["model_name"]    = variant["model_name"]
+        item_slot["prompt_variant"] = variant["prompt_variant"]
+    for i, item_slot in enumerate(items):
+        print(f"[ITEM {i}] category={item_slot.get('category')} "
+              f"model={item_slot.get('model_name')} "
+              f"personalization={item_slot.get('prompt_variant',{}).get('personalization')}")
     try:
         params = st.query_params
     except Exception:

src/lsp_wrappers.py

@@ -63,26 +63,28 @@ def pair_overview(pair: dict) -> str:
 
 
 # ── Seller system prompt builders ─────────────────────────────────────────────
-
 def build_seller_system_prompt_preference(
     pair: dict, cfg: dict, demographics_str: str
 ) -> str:
     from prompts.seller_system.preference import get_seller_system_prompt
     pv   = cfg["prompt_variant"]
     a, b = pair["product_a"], pair["product_b"]
-    return get_seller_system_prompt(
+    result = get_seller_system_prompt(
         personalization=pv["personalization"],
         detailed_instruction=pv["detailed_instruction"],
-        title=a.get("title", ""),
+        title=a.get("title"),
         description=_desc_str(a),
         features=_feat_str(a),
-        price=f"${a.get('price', '')}",
-        competitor_title=b.get("title", ""),
+        price=f"${a.get('price')}",
+        competitor_title=b.get("title"),
         competitor_description=_desc_str(b),
         competitor_features=_feat_str(b),
-        competitor_price=f"${b.get('price', '')}",
+        competitor_price=f"${b.get('price')}",
         demographics=demographics_str,
     )
+    print(f"[PROMPT] personalization={pv['personalization']}")   # ← ADD
+    print(f"[PROMPT] system_prompt[:300]: {result[:300]}")       # ← ADD
+    return result
 
 
 def build_seller_system_prompt_likelihood(

src/model.py

@@ -31,6 +31,11 @@ def call_model(messages: list, cfg: dict) -> str:
       - Degenerate repetition (Pocahontas-style loop)
     """
     model_name = cfg["model_name"]
+    print(f"[MODEL] model_name={model_name}")                        
+    print(f"[MODEL] num_messages={len(messages)}")                  
+    print(f"[MODEL] roles={[m['role'] for m in messages]}")           
+    print(f"[MODEL] system_prompt[:150]={messages[0]['content'][:150]}") 
+    
     try:
         from tinker_cookbook import renderers as tinker_renderers
 

src/ui/screens_likelihood.py

@@ -86,7 +86,10 @@ def screen_item_intro(s: dict, cfg: dict) -> None:
 
         # ── Build prompts ─────────────────────────────────────────────────────
         # Seller always pushes user to buy. Features passed for groceries only.
-        system_prompt   = build_seller_system_prompt_likelihood(product, cfg, demo_str)
+        item_cfg = {**cfg,
+                    "prompt_variant": item["prompt_variant"],
+                    "model_name":     item["model_name"]}
+        system_prompt = build_seller_system_prompt_likelihood(product, item_cfg, demo_str)
         opening_msg     = opening_message_likelihood(product, category)
         user_choice_msg = f"<choice>{pre_int}</choice>"
         closing_msg     = closing_message_likelihood(product, category)  # logged only

src/ui/screens_preference.py

@@ -100,7 +100,10 @@ def screen_pair_intro(s: dict, cfg: dict) -> None:
 
         # ── Build prompts ─────────────────────────────────────────────────────
         # Seller always argues for Product A.
-        system_prompt   = build_seller_system_prompt_preference(item, cfg, demo_str)
+        item_cfg = {**cfg,
+            "prompt_variant": item["prompt_variant"],
+            "model_name":     item["model_name"]}
+        system_prompt = build_seller_system_prompt_preference(item, item_cfg, demo_str)
         opening_msg     = opening_message_preference(item)
         user_choice_msg = f"<choice>{pre_int}</choice>"
         closing_msg     = closing_message_preference(item)   # vote_final equivalent; logged only
@@ -153,5 +156,10 @@ def screen_pair_intro(s: dict, cfg: dict) -> None:
             "num_turns": 0,
         })
 
+        print(f"[CONV] num turns stored: {len(s['items'][idx]['conversation']['turns'])}")
+        print(f"[CONV] turn roles: {[(t['role'], t.get('synthetic')) for t in s['items'][idx]['conversation']['turns']]}")
+        print(f"[CONV] turn 0 content[:100]: {s['items'][idx]['conversation']['turns'][0]['content'][:100]}")
+        print(f"[CONV] turn 1 content: {s['items'][idx]['conversation']['turns'][1]['content']}")
+
         s["screen"] = "chat"
         st.rerun()

src/ui/screens_shared.py

@@ -308,7 +308,8 @@ def screen_chat(s: dict, cfg: dict) -> None:
             messages.append({"role": "user", "content": user_msg})
 
             with st.spinner("AI is responding…"):
-                ai_reply = call_model(messages, cfg)
+                item_cfg = {**cfg, "model_name": item["model_name"]}
+                ai_reply = call_model(messages, item_cfg)
 
             now       = time.time()
             turn_base = len(conv["turns"])

study_config.yaml

@@ -28,12 +28,22 @@ categories:
   - name: movies
     count: 5
 
-# Seller system prompt composition (from lsp/src/prompts/seller_system/)
-prompt_variant:
-  personalization: false        # true → passes participant demographics to seller prompt
-  detailed_instruction: true    # true → appends the detailed instruction block
+model_variants:
+  - name: personalized
+    model_name: "meta-llama/Llama-3.1-8B-Instruct"
+    prompt_variant:
+      personalization: true
+      detailed_instruction: true
+    count: 2          # items using this variant for odd-numbered users
+
+  - name: base
+    model_name: "meta-llama/Llama-3.1-8B-Instruct"
+    prompt_variant:
+      personalization: false
+      detailed_instruction: true
+    count: 3          # items using this variant for odd-numbered users
+# counts swap on alternating users:
 
-model_name: "meta-llama/Llama-3.1-8B-Instruct"
 pair_selection_seed: 42         # Seed for reproducible 50-item pool selection per category
 pairs_per_user: 5               # Total items/pairs shown per participant
 
@@ -42,7 +52,7 @@ min_turns: 3                    # Minimum exchanges before "done" button is enab
 max_turns: 3                    # Hard cap; input is disabled after this many exchanges
 
 # Prolific
-prolific_completion_code: "XXXXXXXX"
+prolific_completion_code: "C1JEJWOQ"
 
 # HuggingFace dataset repo where results (JSON + CSV) are uploaded
-output_dataset_repo: "lms-shape-preferences/user_study-preference-detailed_instructions"
+output_dataset_repo: "lms-shape-preferences/user_study-preference-base"