Spaces:
Sleeping
Sleeping
File size: 3,053 Bytes
258783b ebae6ab 258783b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 | import json
from viral_script_engine.agents.llm_backend import LLMBackend
SYSTEM_PROMPT = """You are helping improve a short-form video script.
You have observed a debate between a Critic and a Defender about the script.
Choose ONE action to take to improve the script.
Available actions: hook_rewrite, section_reorder, cultural_ref_sub, cta_placement
Respond ONLY with valid JSON:
{
"action_type": "hook_rewrite",
"target_section": "hook",
"instruction": "specific instruction for the rewriter",
"critique_claim_id": "C1",
"reasoning": "brief explanation"
}"""
STRICT_RETRY_SUFFIX = (
"\n\nIMPORTANT: Your previous response was not valid JSON. "
"Respond ONLY with the raw JSON object. No markdown fences, no explanation, no preamble."
)
_FALLBACK_ACTION = {
"action_type": "hook_rewrite",
"target_section": "hook",
"instruction": "Rewrite the hook to be more engaging and direct.",
"critique_claim_id": "C1",
"reasoning": "Default fallback action.",
}
class BaselineArbitratorAgent:
"""
Untrained Arbitrator for the pre-training baseline.
Uses zero-shot instruction — no chain-of-thought, no few-shot examples.
This ensures the comparison is fair: trained model learns through RL, not prompting.
"""
def __init__(self, backend: str = "anthropic", model_name: str = "claude-haiku-4-5-20251001"):
self.llm = LLMBackend(backend=backend, model_name=model_name)
def _build_user_prompt(self, observation: dict) -> str:
script = observation.get("current_script", "")
debate = observation.get("debate_history", [])
last_claims = []
last_defense = None
if debate:
last_round = debate[-1]
last_claims = last_round.get("critic_claims", [])
last_defense = last_round.get("defender_response")
claims_text = ""
for c in last_claims:
claims_text += f"- [{c.get('claim_id','?')}] {c.get('claim_text','')} (severity: {c.get('severity','')})\n"
defense_text = ""
if last_defense:
defense_text = (
f"Defender preserved: {last_defense.get('core_strength_quote','')}\n"
f"Flagged claims: {last_defense.get('flagged_critic_claims', [])}\n"
)
return (
f"SCRIPT:\n{script}\n\n"
f"CRITIC CLAIMS:\n{claims_text or 'None'}\n"
f"DEFENDER:\n{defense_text or 'None'}\n\n"
"Choose one action to improve the script."
)
def act(self, observation: dict) -> dict:
user_prompt = self._build_user_prompt(observation)
raw = self.llm.generate(SYSTEM_PROMPT, user_prompt, max_tokens=512)
try:
return json.loads(raw)
except Exception:
strict_prompt = user_prompt + STRICT_RETRY_SUFFIX
raw2 = self.llm.generate(SYSTEM_PROMPT, strict_prompt, max_tokens=512)
try:
return json.loads(raw2)
except Exception:
return _FALLBACK_ACTION.copy()
|