File size: 9,142 Bytes
fe406e9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
"""
Baseline inference script for IndicScriptureQA.

Runs an LLM agent against all 3 tasks via the OpenEnv HTTP API.
Emits structured [START]/[STEP]/[END] logs per the OpenEnv spec.

The agent evaluates BOTH factual accuracy AND semantic structure:
  - factual: hallucination detection, correction
  - structural: coherence, completeness, terminology, logical ordering

Environment variables:
  API_BASE_URL   LLM endpoint         (default: HF router)
  MODEL_NAME     Model identifier      (default: Qwen2.5-72B-Instruct)
  HF_TOKEN       API key
  PING_URL       Environment server    (default: http://localhost:8000)
"""

import json
import os
import textwrap
from typing import Dict, List, Optional

import requests
from openai import OpenAI

# ── Config ────────────────────────────────────────────────────────────────────

API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
ENV_URL = os.getenv("PING_URL", "http://localhost:8000")
BENCHMARK = "indic_scripture_qa"
TEMPERATURE = 0.4
MAX_TOKENS = 600

TASKS = [
    {"name": "verify-factual",    "max_steps": 5},
    {"name": "correct-and-cite",  "max_steps": 8},
    {"name": "fix-hallucination", "max_steps": 12},
]

SYSTEM_PROMPT = textwrap.dedent("""\
You are an expert agent that both CORRECTS hallucinations and IMPROVES the
semantic structure of answers about Indic scriptures (Vedas, Upanishads,
Ramayana, Mahabharata, Bhagavad Gita, Puranas).

Each turn you receive an observation with:
  - question, current_answer, retrieved_passages, current_citations,
    steps_remaining, feedback, structural_hints

You must reply with EXACTLY ONE JSON object (no markdown, no explanation):
{
  "action_type": "RETRIEVE" | "EDIT" | "RESTRUCTURE" | "CITE" | "ACCEPT" | "REJECT",
  "payload": "<string or null>"
}

Actions:
  RETRIEVE    β€” fetch source passages to verify facts
  EDIT        β€” rewrite the answer to fix factual errors AND improve content
  RESTRUCTURE β€” reorganise the answer's flow, ordering, and coherence WITHOUT
                changing facts (use when facts are right but structure is poor)
  CITE        β€” add a scripture citation (e.g. "Bhagavad Gita 2.47")
  ACCEPT      β€” finalise when answer is both accurate and well-structured
  REJECT      β€” only if the answer is fundamentally unsalvageable

Strategy:
  1. RETRIEVE first (1–2 times) to get authoritative source passages.
  2. Check facts against retrieved passages. EDIT to fix any errors.
  3. Read structural_hints. If the answer's flow, terminology, or completeness
     is poor, use RESTRUCTURE to reorganise it.
  4. CITE relevant scripture references.
  5. ACCEPT when the answer is factually accurate, well-structured, uses
     correct Sanskrit terminology, and covers all required aspects.
  6. Be efficient β€” fewer steps score higher.

Evaluation axes (the grader checks ALL of these):
  - Factual similarity to ground truth
  - Citation accuracy
  - Terminology precision (correct Sanskrit/domain terms, no misconception markers)
  - Completeness (all required conceptual aspects covered)
  - Logical ordering (concepts in proper sequence)
  - Coherence (smooth transitions, balanced sentence structure)
""")


# ── Logging helpers ───────────────────────────────────────────────────────────

def log_start(task: str, env: str, model: str) -> None:
    print(f"[START] task={task} env={env} model={model}", flush=True)


def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:
    print(
        f"[STEP] step={step} action={action} reward={reward:.2f} "
        f"done={str(done).lower()} error={error or 'null'}",
        flush=True,
    )


def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
    rewards_str = ",".join(f"{r:.2f}" for r in rewards)
    print(
        f"[END] success={str(success).lower()} steps={steps} "
        f"score={score:.2f} rewards={rewards_str}",
        flush=True,
    )


# ── Env HTTP helpers ──────────────────────────────────────────────────────────

def env_reset(task_name: str, scenario_index: int = 0) -> Dict:
    resp = requests.post(
        f"{ENV_URL}/reset",
        json={"task_name": task_name, "scenario_index": scenario_index},
        timeout=30,
    )
    resp.raise_for_status()
    return resp.json()


def env_step(action_type: str, payload: Optional[str] = None) -> Dict:
    resp = requests.post(
        f"{ENV_URL}/step",
        json={"action_type": action_type, "payload": payload},
        timeout=30,
    )
    resp.raise_for_status()
    return resp.json()


# ── Agent ─────────────────────────────────────────────────────────────────────

def build_user_prompt(obs: Dict, step: int) -> str:
    return json.dumps({
        "step": step,
        "question": obs["question"],
        "current_answer": obs["current_answer"],
        "retrieved_passages": obs["retrieved_passages"],
        "current_citations": obs["current_citations"],
        "steps_remaining": obs["steps_remaining"],
        "feedback": obs.get("feedback"),
        "structural_hints": obs.get("structural_hints", []),
    }, indent=2)


def get_agent_action(client: OpenAI, obs: Dict, step: int) -> Dict:
    """Ask the LLM for the next action."""
    user_prompt = build_user_prompt(obs, step)
    try:
        completion = client.chat.completions.create(
            model=MODEL_NAME,
            messages=[
                {"role": "system", "content": SYSTEM_PROMPT},
                {"role": "user", "content": user_prompt},
            ],
            temperature=TEMPERATURE,
            max_tokens=MAX_TOKENS,
            stream=False,
        )
        raw = (completion.choices[0].message.content or "").strip()
        if raw.startswith("```"):
            raw = raw.split("\n", 1)[-1].rsplit("```", 1)[0].strip()
        return json.loads(raw)
    except Exception as exc:
        print(f"[DEBUG] LLM parse error: {exc}", flush=True)
        if step <= 2:
            return {"action_type": "RETRIEVE", "payload": None}
        return {"action_type": "ACCEPT", "payload": None}


# ── Main loop ─────────────────────────────────────────────────────────────────

def run_task(client: OpenAI, task_name: str, max_steps: int, scenario_index: int = 0) -> float:
    """Run one episode. Returns score in [0, 1]."""
    log_start(task=task_name, env=BENCHMARK, model=MODEL_NAME)

    rewards: List[float] = []
    steps_taken = 0
    score = 0.0
    success = False

    try:
        result = env_reset(task_name, scenario_index)
        obs = result["observation"]

        for step in range(1, max_steps + 1):
            if result.get("done", False):
                break

            agent_action = get_agent_action(client, obs, step)
            action_type = agent_action.get("action_type", "ACCEPT")
            payload = agent_action.get("payload")

            result = env_step(action_type, payload)
            obs = result["observation"]
            reward = result.get("reward", 0.0)
            done = result.get("done", False)

            rewards.append(reward)
            steps_taken = step

            action_str = f"{action_type}({payload!r})" if payload else action_type
            log_step(step=step, action=action_str, reward=reward, done=done, error=None)

            if done:
                score = result.get("info", {}).get("score", 0.0)
                break

        success = score >= 0.10

    except Exception as exc:
        print(f"[DEBUG] Episode error: {exc}", flush=True)

    finally:
        log_end(success=success, steps=steps_taken, score=score, rewards=rewards)

    return score


def main() -> None:
    client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)

    all_scores: Dict[str, float] = {}

    for task in TASKS:
        task_name = task["name"]
        max_steps = task["max_steps"]
        score = run_task(client, task_name, max_steps, scenario_index=0)
        all_scores[task_name] = score
        print(flush=True)

    print("=" * 60, flush=True)
    print("BASELINE RESULTS", flush=True)
    for name, sc in all_scores.items():
        print(f"  {name:25s}  score={sc:.3f}", flush=True)
    avg = sum(all_scores.values()) / len(all_scores) if all_scores else 0.0
    print(f"  {'AVERAGE':25s}  score={avg:.3f}", flush=True)
    print("=" * 60, flush=True)


if __name__ == "__main__":
    main()