Remove natural language input mode and simplify prompt building

app.py

@@ -3,22 +3,10 @@ from pydantic import BaseModel
 from typing import Any, Dict, List, Optional
 
 import json
-import sys
-from pathlib import Path
-
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from peft import PeftModel
 
-ROOT_DIR = Path(__file__).resolve().parent
-RL_DIR = ROOT_DIR / "RL"
-for path in (ROOT_DIR, RL_DIR):
-    path_str = str(path)
-    if path_str not in sys.path:
-        sys.path.append(path_str)
-
-from RL.battleground_nl_utils import game_state_to_natural_language
-
 
 BASE_MODEL_ID = "Qwen/Qwen3-4B-Instruct-2507"
 ADAPTER_MODEL_ID = "iteratehack/battleground-rlaif-qwen-gamehistory-grpo"
@@ -50,55 +38,30 @@ Rules:
 Now here is the game state JSON:
 """
 
-INSTRUCTION_PREFIX_NL = """You are a Hearthstone Battlegrounds AI.
-Given the following natural language description of the current game state, choose
-the best full-turn sequence of actions and respond with a single JSON object in
-this exact format:
-{"actions":[{"type":"<ACTION_TYPE>","tavern_index":<int-or-null>,"hand_index":<int-or-null>,"board_index":<int-or-null>,"card_name":<string-or-null>}, ...]}
-Rules:
-1. Respond with JSON only. Do not add explanations or any extra text.
-2. The top-level object must have exactly one key: "actions".
-3. "actions" must be a JSON array (possibly empty, but usually 1+ steps) of
-   atomic action objects.
-4. Use 0-based integers for indices or null when not used.
-5. "type" must be one of: "BUY_FROM_TAVERN","PLAY_FROM_HAND","SELL_FROM_BOARD",
-   "HERO_POWER","ROLL","UPGRADE_TAVERN","FREEZE","END_TURN".
-6. "card_name" must exactly match a card name from the game state when required,
-   otherwise null.
-Now here is the description of the game state:
-"""
 
 
 class GenerateRequest(BaseModel):
     phase: Optional[str] = None
     turn: Optional[int] = None
     state: Dict[str, Any]
-    input_mode: str = "json"  # "json" or "nl"
     max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS
     temperature: float = DEFAULT_TEMPERATURE
 
 
-def build_prompt(example: Dict[str, Any], input_mode: str = "json") -> str:
+def build_prompt(example: Dict[str, Any]) -> str:
+    """Build a JSON-mode prompt (the only mode supported by this Space)."""
     state = example.get("state", {}) or {}
-
-    if input_mode == "nl":
-        nl_state = game_state_to_natural_language(state)
-        prefix = INSTRUCTION_PREFIX_NL
-        state_text = nl_state
-    else:
-        gs = state.get("game_state", {}) or {}
-        phase = example.get("phase", gs.get("phase", "PlayerTurn"))
-        turn = example.get("turn", gs.get("turn_number", 0))
-        obj = {
-            "task": "battlegrounds_policy_v1",
-            "phase": phase,
-            "turn": turn,
-            "state": state,
-        }
-        state_text = json.dumps(obj, separators=(",", ":"), ensure_ascii=False)
-        prefix = INSTRUCTION_PREFIX
-
-    return prefix + "\n" + state_text
+    gs = state.get("game_state", {}) or {}
+    phase = example.get("phase", gs.get("phase", "PlayerTurn"))
+    turn = example.get("turn", gs.get("turn_number", 0))
+    obj = {
+        "task": "battlegrounds_policy_v1",
+        "phase": phase,
+        "turn": turn,
+        "state": state,
+    }
+    state_text = json.dumps(obj, separators=(",", ":"), ensure_ascii=False)
+    return INSTRUCTION_PREFIX + "\n" + state_text
 
 
 def parse_actions_from_completion(text: str) -> Optional[List[Dict[str, Any]]]:
@@ -199,7 +162,7 @@ def generate_actions(req: GenerateRequest):
         "turn": req.turn,
         "state": req.state,
     }
-    prompt = build_prompt(example, input_mode=req.input_mode)
+    prompt = build_prompt(example)
 
     inputs = tokenizer(prompt, return_tensors="pt")
     inputs = {k: v.to(device) for k, v in inputs.items()}