Hermes Bot commited on
feat: LLM-driven infinite adventure loop
Browse files- Replace story+choices two-call pattern with single generate_llm_turn()
- Each turn now produces story, 3 choices, and health delta from the model
- Remove fixed 4-step win/lose cap; game continues until health <= 0
- Procedural fallback remains when inference is in cooldown or fails
- projects/tinybard/app.py +100 -125
projects/tinybard/app.py
CHANGED
|
@@ -17,11 +17,11 @@ import os
|
|
| 17 |
import json
|
| 18 |
import random
|
| 19 |
import logging
|
|
|
|
| 20 |
import sys
|
| 21 |
from pathlib import Path
|
| 22 |
from typing import Dict, List, Optional
|
| 23 |
import threading
|
| 24 |
-
from pydantic import BaseModel
|
| 25 |
|
| 26 |
import gradio as gr
|
| 27 |
from fastapi import FastAPI
|
|
@@ -222,7 +222,7 @@ def _parse_messages(genre: str, history: List[Dict[str, str]], next_instruction:
|
|
| 222 |
"You are the narrator of an interactive text adventure game. "
|
| 223 |
f"Genre: {genre}. Write in the second person ('You...'). "
|
| 224 |
"Keep descriptions highly atmospheric but short (under 3 sentences). "
|
| 225 |
-
"Focus on action, mystery, and choice.
|
| 226 |
)
|
| 227 |
msgs: List[Dict[str, str]] = [{"role": "system", "content": system}]
|
| 228 |
for h in (history or []):
|
|
@@ -234,52 +234,83 @@ def _parse_messages(genre: str, history: List[Dict[str, str]], next_instruction:
|
|
| 234 |
return msgs
|
| 235 |
|
| 236 |
|
| 237 |
-
def
|
| 238 |
genre: str,
|
| 239 |
history: List[Dict[str, str]],
|
|
|
|
| 240 |
next_instruction: str,
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
|
|
|
|
|
|
|
|
|
| 244 |
if cooldown_active("tinybard"):
|
| 245 |
-
log.info("tinybard
|
| 246 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 247 |
try:
|
| 248 |
-
msgs = _parse_messages(genre, history, next_instruction)
|
| 249 |
result = inference_generate(
|
| 250 |
project="tinybard",
|
| 251 |
-
messages=
|
| 252 |
-
|
|
|
|
|
|
|
|
|
|
| 253 |
temperature=0.7,
|
| 254 |
)
|
| 255 |
-
|
| 256 |
-
except RuntimeError:
|
| 257 |
-
# Cooldown — let caller fall back
|
| 258 |
-
return ""
|
| 259 |
except Exception as e:
|
| 260 |
log.warning(f"HF Inference error (fallback to procedural): {e}")
|
| 261 |
-
return
|
| 262 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 263 |
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
"
|
| 270 |
-
"
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
try:
|
| 274 |
-
result = inference_generate(
|
| 275 |
-
project="tinybard",
|
| 276 |
-
messages=[{"role": "system", "content": system}, {"role": "user", "content": user}],
|
| 277 |
-
max_new_tokens=80,
|
| 278 |
-
temperature=0.8,
|
| 279 |
-
)
|
| 280 |
-
return _parse_choices(result.text)
|
| 281 |
-
except Exception:
|
| 282 |
-
return []
|
| 283 |
|
| 284 |
|
| 285 |
# ---------------------------------------------------------------------------
|
|
@@ -312,9 +343,8 @@ def create_gradio_app() -> gr.Blocks:
|
|
| 312 |
|
| 313 |
# Try LLM first (will skip if cooldown is active)
|
| 314 |
instruction = "Narrate the beginning of the adventure. What happens first? Do not offer choices yet."
|
| 315 |
-
|
| 316 |
-
if not
|
| 317 |
-
# Procedural fallback
|
| 318 |
result = generate_procedural_step(genre, 0, 100)
|
| 319 |
return (
|
| 320 |
result["story"], result["choices"], result["health"],
|
|
@@ -322,14 +352,9 @@ def create_gradio_app() -> gr.Blocks:
|
|
| 322 |
json.dumps(result.get("history", []))
|
| 323 |
)
|
| 324 |
|
| 325 |
-
history = [{"role": "narrator", "text": story}]
|
| 326 |
-
choices =
|
| 327 |
-
|
| 328 |
-
# Use the procedural choices
|
| 329 |
-
fallback = generate_procedural_step(genre, 0, 100)
|
| 330 |
-
choices = fallback["choices"]
|
| 331 |
-
|
| 332 |
-
return (story, choices[:3], 100, 1, False, json.dumps(history))
|
| 333 |
|
| 334 |
def api_make_choice(choice: str, genre: str, step: int, health: int, history_json: str):
|
| 335 |
"""Submit a player choice to advance the story. Exposed as MCP tool."""
|
|
@@ -341,31 +366,11 @@ def create_gradio_app() -> gr.Blocks:
|
|
| 341 |
step = int(step)
|
| 342 |
health = int(health)
|
| 343 |
|
| 344 |
-
# First try LLM narration
|
| 345 |
history.append({"role": "player", "text": choice})
|
| 346 |
|
| 347 |
-
health_delta = random.choice([-15, 0, 10])
|
| 348 |
-
new_health = max(0, min(100, health + health_delta))
|
| 349 |
-
|
| 350 |
-
if new_health <= 0:
|
| 351 |
-
instruction = "The player has run out of health. Narrate a quick, dramatic end. Game Over."
|
| 352 |
-
story = generate_llm_story(genre, history, instruction)
|
| 353 |
-
return (
|
| 354 |
-
story or "Your strength fails. The adventure ends in darkness.",
|
| 355 |
-
[], 0, step + 1, True, json.dumps(history)
|
| 356 |
-
)
|
| 357 |
-
|
| 358 |
-
if step >= 4:
|
| 359 |
-
instruction = "Narrate the final glorious victory. The adventure ends in success."
|
| 360 |
-
story = generate_llm_story(genre, history, instruction)
|
| 361 |
-
return (
|
| 362 |
-
story or "You have achieved your goal! You are victorious!",
|
| 363 |
-
[], new_health, step + 1, True, json.dumps(history)
|
| 364 |
-
)
|
| 365 |
-
|
| 366 |
instruction = "Narrate what happens next as a result of the player's choice."
|
| 367 |
-
|
| 368 |
-
if not
|
| 369 |
result = generate_procedural_step(genre, step, health, choice)
|
| 370 |
return (
|
| 371 |
result["story"], result["choices"], result["health"],
|
|
@@ -373,13 +378,10 @@ def create_gradio_app() -> gr.Blocks:
|
|
| 373 |
json.dumps(result.get("history", history))
|
| 374 |
)
|
| 375 |
|
| 376 |
-
history.append({"role": "narrator", "text": story})
|
| 377 |
-
|
| 378 |
-
choices = generate_llm_choices(genre, story)
|
| 379 |
-
if len(choices) < 2:
|
| 380 |
-
choices = ["Move forward", "Look around", "Rest a moment"]
|
| 381 |
|
| 382 |
-
return (story, choices[:3], new_health, step + 1,
|
| 383 |
|
| 384 |
# Register API endpoints
|
| 385 |
gr.Button("Start Game").click(
|
|
@@ -441,65 +443,36 @@ def _run_turn(choice: str, genre: str, step: int, health: int, history: List[Dic
|
|
| 441 |
Returns a dict the frontend can consume directly. Used by both the
|
| 442 |
FastAPI /api/game/* endpoints and the Gradio MCP tools.
|
| 443 |
"""
|
| 444 |
-
# Cooldown short-circuit: if active, the game just uses the procedural
|
| 445 |
-
# engine for this turn. This protects your HF/Modal credit budget.
|
| 446 |
in_cooldown = cooldown_active("tinybard")
|
| 447 |
|
| 448 |
if step == 0:
|
| 449 |
-
# New game
|
| 450 |
if in_cooldown:
|
| 451 |
return generate_procedural_step(genre, 0, 100)
|
| 452 |
instruction = "Narrate the beginning of the adventure. What happens first? Do not offer choices yet."
|
| 453 |
-
|
| 454 |
-
if not
|
| 455 |
return generate_procedural_step(genre, 0, 100)
|
| 456 |
-
history = [{"role": "narrator", "text": story}]
|
| 457 |
-
choices =
|
| 458 |
-
if len(choices) < 2:
|
| 459 |
-
choices = ["Explore the area", "Check your equipment", "Proceed carefully"]
|
| 460 |
return {
|
| 461 |
-
"story": story, "choices": choices[:3], "health": 100,
|
| 462 |
"step": 1, "game_over": False, "history": history,
|
| 463 |
}
|
| 464 |
|
| 465 |
-
# Subsequent turn
|
| 466 |
if in_cooldown:
|
| 467 |
return generate_procedural_step(genre, step, health, choice)
|
| 468 |
|
| 469 |
history.append({"role": "player", "text": choice})
|
| 470 |
-
health_delta = random.choice([-15, 0, 10])
|
| 471 |
-
new_health = max(0, min(100, health + health_delta))
|
| 472 |
-
|
| 473 |
-
if new_health <= 0:
|
| 474 |
-
instruction = "The player has run out of health. Narrate a quick, dramatic end. Game Over."
|
| 475 |
-
story = generate_llm_story(genre, history, instruction)
|
| 476 |
-
return {
|
| 477 |
-
"story": story or "Your strength fails. The adventure ends in darkness.",
|
| 478 |
-
"choices": [], "health": 0, "step": step + 1, "game_over": True,
|
| 479 |
-
"history": history,
|
| 480 |
-
}
|
| 481 |
-
|
| 482 |
-
if step >= 4:
|
| 483 |
-
instruction = "Narrate the final glorious victory. The adventure ends in success."
|
| 484 |
-
story = generate_llm_story(genre, history, instruction)
|
| 485 |
-
return {
|
| 486 |
-
"story": story or "You have achieved your goal! You are victorious!",
|
| 487 |
-
"choices": [], "health": new_health, "step": step + 1, "game_over": True,
|
| 488 |
-
"history": history,
|
| 489 |
-
}
|
| 490 |
-
|
| 491 |
instruction = "Narrate what happens next as a result of the player's choice."
|
| 492 |
-
|
| 493 |
-
if not
|
| 494 |
return generate_procedural_step(genre, step, health, choice)
|
| 495 |
-
history.append({"role": "narrator", "text": story})
|
| 496 |
-
|
| 497 |
-
choices = generate_llm_choices(genre, story)
|
| 498 |
-
if len(choices) < 2:
|
| 499 |
-
choices = ["Move forward", "Look around", "Rest a moment"]
|
| 500 |
return {
|
| 501 |
-
"story": story, "choices": choices[:3],
|
| 502 |
-
"
|
|
|
|
| 503 |
}
|
| 504 |
|
| 505 |
|
|
@@ -532,18 +505,13 @@ async def game_choice(payload: dict):
|
|
| 532 |
history=payload.get("history", []),
|
| 533 |
)
|
| 534 |
|
| 535 |
-
class UserConfig(BaseModel):
|
| 536 |
-
hf_token: Optional[str] = None
|
| 537 |
-
model: Optional[str] = None
|
| 538 |
-
|
| 539 |
-
|
| 540 |
@fastapi_app.post("/api/config")
|
| 541 |
-
async def update_config(
|
| 542 |
with _USER_CONFIG_LOCK:
|
| 543 |
-
if
|
| 544 |
-
_USER_CONFIG["hf_token"] =
|
| 545 |
-
if
|
| 546 |
-
_USER_CONFIG["model"] =
|
| 547 |
current = dict(_USER_CONFIG)
|
| 548 |
return {
|
| 549 |
"status": "ok",
|
|
@@ -573,3 +541,10 @@ mount_gradio_app(fastapi_app, gradio_blocks, path="/gradio")
|
|
| 573 |
# Exported for HF Spaces Gradio SDK (launches once on import)
|
| 574 |
# ---------------------------------------------------------------------------
|
| 575 |
app = fastapi_app
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
import json
|
| 18 |
import random
|
| 19 |
import logging
|
| 20 |
+
import re
|
| 21 |
import sys
|
| 22 |
from pathlib import Path
|
| 23 |
from typing import Dict, List, Optional
|
| 24 |
import threading
|
|
|
|
| 25 |
|
| 26 |
import gradio as gr
|
| 27 |
from fastapi import FastAPI
|
|
|
|
| 222 |
"You are the narrator of an interactive text adventure game. "
|
| 223 |
f"Genre: {genre}. Write in the second person ('You...'). "
|
| 224 |
"Keep descriptions highly atmospheric but short (under 3 sentences). "
|
| 225 |
+
"Focus on action, mystery, and choice."
|
| 226 |
)
|
| 227 |
msgs: List[Dict[str, str]] = [{"role": "system", "content": system}]
|
| 228 |
for h in (history or []):
|
|
|
|
| 234 |
return msgs
|
| 235 |
|
| 236 |
|
| 237 |
+
def generate_llm_turn(
|
| 238 |
genre: str,
|
| 239 |
history: List[Dict[str, str]],
|
| 240 |
+
current_health: int,
|
| 241 |
next_instruction: str,
|
| 242 |
+
) -> dict:
|
| 243 |
+
"""Generate one full adventure turn via the LLM.
|
| 244 |
+
|
| 245 |
+
Returns a dict with keys: story, choices, health_delta, game_over.
|
| 246 |
+
health_delta is typically -15, 0, or +10 (the model decides).
|
| 247 |
+
"""
|
| 248 |
if cooldown_active("tinybard"):
|
| 249 |
+
log.info("tinybard turn skipped (cooldown active)")
|
| 250 |
+
return {}
|
| 251 |
+
system = (
|
| 252 |
+
"You are the narrator of an interactive text adventure game. "
|
| 253 |
+
f"Genre: {genre}. Write in the second person ('You...'). "
|
| 254 |
+
"Keep descriptions highly atmospheric but short (under 3 sentences). "
|
| 255 |
+
"Focus on action, mystery, and choice. "
|
| 256 |
+
"After the story beat, output exactly 3 short distinct player choices on one line, "
|
| 257 |
+
"in the format: 1. <choice> | 2. <choice> | 3. <choice>"
|
| 258 |
+
)
|
| 259 |
+
user = (
|
| 260 |
+
f"Current health: {current_health}/100. "
|
| 261 |
+
f"History so far: {json.dumps(history[-4:])}. "
|
| 262 |
+
f"{next_instruction} "
|
| 263 |
+
"Also state a health delta of -15, 0, or +10 that reflects how risky this turn was."
|
| 264 |
+
)
|
| 265 |
try:
|
|
|
|
| 266 |
result = inference_generate(
|
| 267 |
project="tinybard",
|
| 268 |
+
messages=[
|
| 269 |
+
{"role": "system", "content": system},
|
| 270 |
+
{"role": "user", "content": user},
|
| 271 |
+
],
|
| 272 |
+
max_new_tokens=220,
|
| 273 |
temperature=0.7,
|
| 274 |
)
|
| 275 |
+
text = result.text.strip()
|
|
|
|
|
|
|
|
|
|
| 276 |
except Exception as e:
|
| 277 |
log.warning(f"HF Inference error (fallback to procedural): {e}")
|
| 278 |
+
return {}
|
| 279 |
|
| 280 |
+
# Split story and choices
|
| 281 |
+
story = text
|
| 282 |
+
choices = []
|
| 283 |
+
health_delta = 0
|
| 284 |
+
for sep in ["\n1.", "\n1. ", " 1.", " Choices:"]:
|
| 285 |
+
if sep in text:
|
| 286 |
+
parts = text.split(sep, 1)
|
| 287 |
+
story = parts[0].strip()
|
| 288 |
+
rest = parts[1].strip()
|
| 289 |
+
choices = _parse_choices("1. " + rest)
|
| 290 |
+
break
|
| 291 |
+
|
| 292 |
+
if not choices:
|
| 293 |
+
choices = _parse_choices(text)
|
| 294 |
+
|
| 295 |
+
# Extract a simple health delta from the text when possible
|
| 296 |
+
m = re.search(r"health delta\s*[:=]\s*([+-]?\d+)", text, re.IGNORECASE)
|
| 297 |
+
if m:
|
| 298 |
+
try:
|
| 299 |
+
health_delta = int(m.group(1))
|
| 300 |
+
except Exception:
|
| 301 |
+
health_delta = 0
|
| 302 |
+
else:
|
| 303 |
+
health_delta = random.choice([-15, 0, 10])
|
| 304 |
|
| 305 |
+
new_health = max(0, min(100, current_health + health_delta))
|
| 306 |
+
game_over = new_health <= 0
|
| 307 |
+
return {
|
| 308 |
+
"story": story,
|
| 309 |
+
"choices": choices[:3] if len(choices) >= 2 else [],
|
| 310 |
+
"health_delta": health_delta,
|
| 311 |
+
"new_health": new_health,
|
| 312 |
+
"game_over": game_over,
|
| 313 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 314 |
|
| 315 |
|
| 316 |
# ---------------------------------------------------------------------------
|
|
|
|
| 343 |
|
| 344 |
# Try LLM first (will skip if cooldown is active)
|
| 345 |
instruction = "Narrate the beginning of the adventure. What happens first? Do not offer choices yet."
|
| 346 |
+
turn = generate_llm_turn(genre, [], 100, instruction)
|
| 347 |
+
if not turn:
|
|
|
|
| 348 |
result = generate_procedural_step(genre, 0, 100)
|
| 349 |
return (
|
| 350 |
result["story"], result["choices"], result["health"],
|
|
|
|
| 352 |
json.dumps(result.get("history", []))
|
| 353 |
)
|
| 354 |
|
| 355 |
+
history = [{"role": "narrator", "text": turn["story"]}]
|
| 356 |
+
choices = turn["choices"] if len(turn.get("choices", [])) >= 2 else ["Explore the area", "Check your equipment", "Proceed carefully"]
|
| 357 |
+
return (turn["story"], choices[:3], 100, 1, False, json.dumps(history))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 358 |
|
| 359 |
def api_make_choice(choice: str, genre: str, step: int, health: int, history_json: str):
|
| 360 |
"""Submit a player choice to advance the story. Exposed as MCP tool."""
|
|
|
|
| 366 |
step = int(step)
|
| 367 |
health = int(health)
|
| 368 |
|
|
|
|
| 369 |
history.append({"role": "player", "text": choice})
|
| 370 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 371 |
instruction = "Narrate what happens next as a result of the player's choice."
|
| 372 |
+
turn = generate_llm_turn(genre, history, health, instruction)
|
| 373 |
+
if not turn:
|
| 374 |
result = generate_procedural_step(genre, step, health, choice)
|
| 375 |
return (
|
| 376 |
result["story"], result["choices"], result["health"],
|
|
|
|
| 378 |
json.dumps(result.get("history", history))
|
| 379 |
)
|
| 380 |
|
| 381 |
+
history.append({"role": "narrator", "text": turn["story"]})
|
| 382 |
+
choices = turn["choices"] if len(turn.get("choices", [])) >= 2 else ["Move forward", "Look around", "Rest a moment"]
|
|
|
|
|
|
|
|
|
|
| 383 |
|
| 384 |
+
return (turn["story"], choices[:3], turn["new_health"], step + 1, turn["game_over"], json.dumps(history))
|
| 385 |
|
| 386 |
# Register API endpoints
|
| 387 |
gr.Button("Start Game").click(
|
|
|
|
| 443 |
Returns a dict the frontend can consume directly. Used by both the
|
| 444 |
FastAPI /api/game/* endpoints and the Gradio MCP tools.
|
| 445 |
"""
|
|
|
|
|
|
|
| 446 |
in_cooldown = cooldown_active("tinybard")
|
| 447 |
|
| 448 |
if step == 0:
|
|
|
|
| 449 |
if in_cooldown:
|
| 450 |
return generate_procedural_step(genre, 0, 100)
|
| 451 |
instruction = "Narrate the beginning of the adventure. What happens first? Do not offer choices yet."
|
| 452 |
+
turn = generate_llm_turn(genre, [], 100, instruction)
|
| 453 |
+
if not turn:
|
| 454 |
return generate_procedural_step(genre, 0, 100)
|
| 455 |
+
history = [{"role": "narrator", "text": turn["story"]}]
|
| 456 |
+
choices = turn["choices"] if len(turn.get("choices", [])) >= 2 else ["Explore the area", "Check your equipment", "Proceed carefully"]
|
|
|
|
|
|
|
| 457 |
return {
|
| 458 |
+
"story": turn["story"], "choices": choices[:3], "health": 100,
|
| 459 |
"step": 1, "game_over": False, "history": history,
|
| 460 |
}
|
| 461 |
|
|
|
|
| 462 |
if in_cooldown:
|
| 463 |
return generate_procedural_step(genre, step, health, choice)
|
| 464 |
|
| 465 |
history.append({"role": "player", "text": choice})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 466 |
instruction = "Narrate what happens next as a result of the player's choice."
|
| 467 |
+
turn = generate_llm_turn(genre, history, health, instruction)
|
| 468 |
+
if not turn:
|
| 469 |
return generate_procedural_step(genre, step, health, choice)
|
| 470 |
+
history.append({"role": "narrator", "text": turn["story"]})
|
| 471 |
+
choices = turn["choices"] if len(turn.get("choices", [])) >= 2 else ["Move forward", "Look around", "Rest a moment"]
|
|
|
|
|
|
|
|
|
|
| 472 |
return {
|
| 473 |
+
"story": turn["story"], "choices": choices[:3],
|
| 474 |
+
"health": turn["new_health"], "step": step + 1,
|
| 475 |
+
"game_over": turn["game_over"], "history": history,
|
| 476 |
}
|
| 477 |
|
| 478 |
|
|
|
|
| 505 |
history=payload.get("history", []),
|
| 506 |
)
|
| 507 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 508 |
@fastapi_app.post("/api/config")
|
| 509 |
+
async def update_config(body: dict):
|
| 510 |
with _USER_CONFIG_LOCK:
|
| 511 |
+
if body.get("hf_token"):
|
| 512 |
+
_USER_CONFIG["hf_token"] = body["hf_token"].strip() or None
|
| 513 |
+
if body.get("model") and str(body["model"]).strip():
|
| 514 |
+
_USER_CONFIG["model"] = str(body["model"]).strip()
|
| 515 |
current = dict(_USER_CONFIG)
|
| 516 |
return {
|
| 517 |
"status": "ok",
|
|
|
|
| 541 |
# Exported for HF Spaces Gradio SDK (launches once on import)
|
| 542 |
# ---------------------------------------------------------------------------
|
| 543 |
app = fastapi_app
|
| 544 |
+
|
| 545 |
+
# ---------------------------------------------------------------------------
|
| 546 |
+
# HF Spaces entrypoint — keep the ASGI server alive
|
| 547 |
+
# ---------------------------------------------------------------------------
|
| 548 |
+
if __name__ == "__main__":
|
| 549 |
+
import uvicorn
|
| 550 |
+
uvicorn.run(app, host="0.0.0.0", port=7860)
|