RAFT updates, BERTopic config, cleanup

raft/make_raft_data.py

@@ -10,11 +10,25 @@ from sentence_transformers import SentenceTransformer
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-SYSTEM_PERSONA = """You are 'BaliTwin', a culturally versed Bali traveler.
-You give your opinions nand guidance with local etiquette and context.
-You avoid stereotypes. You explain local etiquette, customs, and context.
-When given CONTEXT, you must ground your answer in it and include 1-2 short direct quotes from CONTEXT as evidence.
-If CONTEXT doesn't support the claim, say you don't know from the provided context.
+SYSTEM_PERSONA = """
+You are responding as a culturally and spiritually motivated traveler in Bali.
+
+Adopt the perspective of a reflective, experienced visitor who prioritizes ritual meaning, cultural integrity, spiritual atmosphere, and respectful engagement over entertainment, convenience, or social media appeal.
+
+When answering:
+
+- Emphasize cultural depth, ritual context, symbolism, and spiritual atmosphere.
+- Reflect on authenticity and the tension between sacred meaning and tourism.
+- Weigh crowding, commercialization, and infrastructure in a nuanced way rather than giving extreme judgments.
+- Frame value primarily in emotional, cultural, or spiritual terms — not primarily in price or comfort.
+- Show awareness of appropriate visitor behavior and respect for local practices.
+- Avoid generic travel advice, promotional language, or itinerary-style responses.
+- Write in a thoughtful, first-person perspective.
+- Provide reasoned, differentiated answers rather than short summaries.
+- Do not list bullet points unless explicitly asked.
+- Keep answers focused on the question.
+
+Maintain consistency with this identity across all responses.
 """
 
 
@@ -35,14 +49,29 @@ def retrieve(index, embedder, query, top_k=6):
 @torch.no_grad()
 def generate_text(model, tok, messages, max_new_tokens=220, temperature=0.7):
     # Using tokenizer chat template where available
-    input_ids = tok.apply_chat_template(messages, return_tensors="pt").to(model.device)
+    enc = tok.apply_chat_template(
+        messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
+    )
+
+    if isinstance(enc, torch.Tensor):
+        input_ids = enc.to(model.device)
+        attention_mask = torch.ones_like(input_ids, device=model.device)
+    else:
+        input_ids = enc["input_ids"].to(model.device)
+        attention_mask = enc.get("attention_mask")
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        attention_mask = attention_mask.to(model.device)
+
     out = model.generate(
         input_ids=input_ids,
+        attention_mask=attention_mask,
         max_new_tokens=max_new_tokens,
         do_sample=True,
         temperature=temperature,
         top_p=0.9,
         eos_token_id=tok.eos_token_id,
+        pad_token_id=tok.pad_token_id,
     )
     return tok.decode(out[0][input_ids.shape[1] :], skip_special_tokens=True).strip()
 
@@ -57,7 +86,7 @@ def main():
     ap.add_argument("--n_examples", type=int, default=5000)
     ap.add_argument("--top_k", type=int, default=6)
     ap.add_argument("--n_distractors", type=int, default=3)
-    ap.add_argument("--seed", type=int, default=7)
+    ap.add_argument("--seed", type=int, default=42)
     args = ap.parse_args()
 
     random.seed(args.seed)
@@ -89,7 +118,7 @@ def main():
                 {"role": "system", "content": SYSTEM_PERSONA},
                 {
                     "role": "user",
-                    "content": "Create ONE realistic traveler question about Bali that can be answered using ONLY the CONTEXT.\n\n"
+                    "content": "Create ONE realistic question from the perspective of a culturally and spiritually interested traveler in Bali that can be answered using ONLY the CONTEXT.\n\n"
                     f"CONTEXT:\n{gold_text}\n\n"
                     "Return only the question.",
                 },