Upload folder using huggingface_hub

train/eval_pivot_set.py

@@ -87,17 +87,20 @@ def main():
                 {"role": "system", "content": SYSTEM_PROMPT},
                 {"role": "user", "content": build_full_observation(clip_id, scenarios)},
             ]
-            inputs = tokenizer.apply_chat_template(
-                messages, return_tensors="pt", add_generation_prompt=True
-            ).to(model.device)
+            encoded = tokenizer.apply_chat_template(
+                messages, return_tensors="pt", add_generation_prompt=True,
+            )
+            input_ids = encoded.input_ids if hasattr(encoded, "input_ids") else encoded
+            input_ids = input_ids.to(model.device)
+            prompt_len = input_ids.shape[1]
             out = model.generate(
-                inputs,
+                input_ids=input_ids,
                 max_new_tokens=args.max_tokens,
                 do_sample=True,
                 temperature=args.temperature,
                 pad_token_id=tokenizer.eos_token_id,
             )
-            text = tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True)
+            text = tokenizer.decode(out[0][prompt_len:], skip_special_tokens=True)
 
             decomp = reward_decomposition(text, gold)
             results.append({

train/sft_warmup.py

@@ -200,14 +200,22 @@ def sample_before_after(model, tokenizer, scenarios, sample_clip_ids, label_for_
             {"role": "system", "content": SYSTEM_PROMPT},
             {"role": "user", "content": build_full_observation(cid, scenarios)},
         ]
-        inputs = tokenizer.apply_chat_template(
-            messages, return_tensors="pt", add_generation_prompt=True
-        ).to(model.device)
+        # apply_chat_template can return either a tensor (older transformers)
+        # or a BatchEncoding (newer transformers). Handle both.
+        encoded = tokenizer.apply_chat_template(
+            messages, return_tensors="pt", add_generation_prompt=True,
+        )
+        if hasattr(encoded, "input_ids"):
+            input_ids = encoded.input_ids.to(model.device)
+        else:
+            input_ids = encoded.to(model.device)
+        prompt_len = input_ids.shape[1]
         out = model.generate(
-            inputs, max_new_tokens=350, do_sample=True, temperature=0.7,
+            input_ids=input_ids,
+            max_new_tokens=350, do_sample=True, temperature=0.7,
             pad_token_id=tokenizer.eos_token_id,
         )
-        text = tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True)
+        text = tokenizer.decode(out[0][prompt_len:], skip_special_tokens=True)
         print(f"\nClip {cid} (gold={gold}, speaker={sc.get('speaker')}):")
         print(text[:1000])
         print("---")

train/side_by_side.py

@@ -95,14 +95,17 @@ def generate_completion(model, tokenizer, prompt_user_msg, max_tokens=600, tempe
         {"role": "system", "content": SYSTEM_PROMPT},
         {"role": "user", "content": prompt_user_msg},
     ]
-    inputs = tokenizer.apply_chat_template(
-        messages, return_tensors="pt", add_generation_prompt=True
-    ).to(model.device)
+    encoded = tokenizer.apply_chat_template(
+        messages, return_tensors="pt", add_generation_prompt=True,
+    )
+    input_ids = encoded.input_ids if hasattr(encoded, "input_ids") else encoded
+    input_ids = input_ids.to(model.device)
+    prompt_len = input_ids.shape[1]
     out = model.generate(
-        inputs, max_new_tokens=max_tokens, do_sample=True,
+        input_ids=input_ids, max_new_tokens=max_tokens, do_sample=True,
         temperature=temperature, pad_token_id=tokenizer.eos_token_id,
     )
-    return tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True)
+    return tokenizer.decode(out[0][prompt_len:], skip_special_tokens=True)
 
 
 def main():