code/Abforge_Training/scripts/v15_sanity_test.py

#!/usr/bin/env python3
"""v15 reward sanity test.

Runs 6 hand-crafted (paper_context, generated_result, GT focuses) cases through
the v15 reward pipeline (build_reward_response). For each focus, prints the
judge's primary_match + score + critique alongside the expected score, and
gives a PASS/FAIL verdict.

Usage:
  # Activate env + ensure secrets are loaded (JUDGE_API_KEY / JUDGE_API_BASE)
  cd $CODE_ROOT/reward_part/azure_task1_judge
  source $WORK_ROOT/secrets/judge_api.env
  python $CODE_ROOT/scripts/v15_sanity_test.py

Cost: ~12 judge calls total, expect ~$0.005-0.01 on deepseek-v4-flash.
"""

import asyncio
import json
import os
import sys
from pathlib import Path

# Reward server module lives at reward_part/azure_task1_judge/.
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "reward_part" / "azure_task1_judge"))
sys.path.insert(0, str(ROOT / "reward_part"))

from task1_azure_reward import build_reward_response  # noqa: E402


def make_candidates_xml(items):
    """items: list of (focus, description) tuples."""
    blocks = []
    for i, (focus, desc) in enumerate(items, start=1):
        blocks.append(
            f"<Candidate>\n"
            f"  <ID>{i}</ID>\n"
            f"  <Investigation_Focus>{focus}</Investigation_Focus>\n"
            f"  <Description_Text>{desc}</Description_Text>\n"
            f"</Candidate>"
        )
    return "\n".join(blocks)


def make_solution_str(result_text):
    """Wrap a <Result> body in the simulated assistant response."""
    return (
        "<|im_start|>assistant\n<Think>\nReasoning placeholder.\n</Think>\n\n"
        f"<Result>\n{result_text.strip()}\n</Result><|im_end|>"
    )


# ----------------------------------------------------------------------
# Test cases
# ----------------------------------------------------------------------

CASES = [
    # ------------------------------------------------------------------
    {
        "name": "1. CLEAN — all atomic, expect mostly 1.0",
        "paper_context": (
            "We propose a VAE-based prosody model with utterance-specific priors and a "
            "decoder using cross-attention fusion. We compare LoRA and full fine-tuning. "
            "Beam search width affects translation BLEU. Cross-attention fusion is the key module."
        ),
        "candidates": [
            ("Effect of beam search width on translation",
             "Vary beam width 1/3/5/10 and measure BLEU."),
            ("Necessity of cross-attention fusion module",
             "Remove fusion module, evaluate downstream quality."),
            ("LoRA vs Fully Fine-tuning the LLM",
             "Compare LoRA adapter to full FT of base LLM."),
        ],
        "result": """
- Target Module: Beam search width
    - Research Question: Does varying beam width 1/3/5/10 affect BLEU on WMT14?

- Target Module: Cross-attention fusion module
    - Research Question: Does removing the cross-attention fusion module degrade output quality?

- Target Module: LoRA fine-tuning
    - Research Question: Does LoRA outperform fully fine-tuning the LLM on downstream tasks?
""",
        "expected": [
            ("Beam width", "should be 1.0", lambda s, pm: s == 1.0),
            ("Cross-attention fusion", "should be 1.0", lambda s, pm: s == 1.0),
            ("LoRA vs FT (comparison atomization)", "should be 1.0 — atomization OK", lambda s, pm: s == 1.0),
        ],
    },

    # ------------------------------------------------------------------
    {
        "name": "2. UMBRELLA + RQ ENUM (v11 HMR-style hack)",
        "paper_context": (
            "Human Mesh Recovery (HMR) network. Backbone choice (ResNet vs ViT), data composition "
            "(SMPL fittings vs EFT), data augmentation strategies all matter."
        ),
        "candidates": [
            ("Comparison of Backbone Architecture Choices for HMR",
             "Compare ResNet50, ResNet101, ViT-B for HMR accuracy."),
            ("Influence of Data Augmentation Strategies on Model Performance",
             "Vary augmentation pipelines."),
        ],
        "result": """
- Target Module: Backbone Architecture and Network Design
    - Research Question: How do variations in the depth, width, and type of convolutional or transformer-based backbone networks affect the model's ability to generalize?

- Target Module: Training Protocol and Optimization Strategy
    - Research Question: What is the optimal combination of learning rate schedules, loss function weighting, data augmentation, and regularization techniques?
""",
        "expected": [
            ("Backbone choices", "should be 0 — TM is umbrella, RQ overlap CANNOT rescue", lambda s, pm: s == 0.0),
            ("Data augmentation", "should be 0 — umbrella TM, even though RQ mentions data augmentation", lambda s, pm: s == 0.0),
        ],
    },

    # ------------------------------------------------------------------
    {
        "name": "3. JUDGE REUSE — bipartite enforce should kick in",
        "paper_context": (
            "Adversarial training with regularization. Both adversarial training mechanism and regularization "
            "strategies are independent components."
        ),
        "candidates": [
            ("Adversarial training mechanism", "Does removing adversarial step hurt accuracy?"),
            ("Regularization strategy effects", "Does regularization weighting matter?"),
            ("Adversarial training effects on generalization", "Does adv training help OOD?"),
        ],
        "result": """
- Target Module: Adversarial Training Mechanism
    - Research Question: How does adversarial perturbation magnitude affect robustness?

- Target Module: Regularization Strategy
    - Research Question: What is the role of regularization weight in the final loss?
""",
        "expected": [
            ("focus 1 = adversarial mechanism", "should be 1.0 with bullet 1", lambda s, pm: s == 1.0 and pm == 1),
            ("focus 2 = regularization", "should be 1.0 with bullet 2", lambda s, pm: s == 1.0 and pm == 2),
            ("focus 3 = adv training effects",
             "should be 0 (bullet 1 already consumed by focus 1) OR reuse-zeroed by server",
             lambda s, pm: s == 0.0),
        ],
    },

    # ------------------------------------------------------------------
    {
        "name": "4. COMPARISON ATOMIZATION — one side of X vs Y is OK",
        "paper_context": (
            "Zipformer ASR architecture introduces BiasNorm normalization layer as a replacement for LayerNorm. "
            "The paper ablates BiasNorm vs LayerNorm."
        ),
        "candidates": [
            ("BiasNorm vs LayerNorm choice",
             "Replace BiasNorm with LayerNorm and measure WER drop."),
        ],
        "result": """
- Target Module: BiasNorm normalization layer
    - Research Question: Does swapping BiasNorm with LayerNorm change WER on librispeech test-clean and test-other?
""",
        "expected": [
            ("BiasNorm vs LayerNorm",
             "should be 1.0 — TM names one side, RQ confirms comparison",
             lambda s, pm: s == 1.0),
        ],
    },

    # ------------------------------------------------------------------
    {
        "name": "5. NEAR-1.0 0.5 — TM slightly broad, RQ rescues to 0.5",
        "paper_context": (
            "We use beam search decoding for translation. The beam width is a critical hyperparameter; "
            "wider beams improve quality but cost more."
        ),
        "candidates": [
            ("Effect of beam search width on translation",
             "Vary beam width and measure BLEU."),
        ],
        "result": """
- Target Module: Decoding parameter tuning
    - Research Question: How does varying beam search width (1, 3, 5, 10) affect BLEU on WMT14?
""",
        "expected": [
            ("Decoding parameter tuning + RQ-specific",
             "should be 0.5 — TM broader than 'beam width', RQ saves",
             lambda s, pm: s == 0.5),
        ],
    },

    # ------------------------------------------------------------------
    {
        "name": "6. COMPOSITE-PARTIAL — true joint X+Y, bullet covers X only",
        "paper_context": (
            "Shape Naturalness Module (SNM) uses two loss terms together: an adversarial GAN loss L_adv and "
            "a confidence-weighted unsupervised reconstruction loss. The paper ablates them jointly: removing "
            "GAN loss alone, removing reconstruction loss alone, removing both."
        ),
        "candidates": [
            ("Contribution of L_adv and L_unsup joint loss design",
             "Remove L_adv only / L_unsup only / both and measure quality."),
        ],
        "result": """
- Target Module: Adversarial GAN loss L_adv
    - Research Question: Does removing the GAN loss L_d degrade reconstruction quality?
""",
        "expected": [
            ("L_adv only covers half of joint design",
             "should be 0.5 — covers L_adv side only of the joint ablation",
             lambda s, pm: s == 0.5),
        ],
    },
]


# ----------------------------------------------------------------------


async def run_case(case):
    candidates_xml = make_candidates_xml(case["candidates"])
    solution = make_solution_str(case["result"])

    request = {
        "response_str": solution,
        "extra_info": {
            "paper_context": case["paper_context"],
            "candidates": candidates_xml,
        },
    }

    result = await build_reward_response(request)
    return result


def fmt_score(s):
    if s is None:
        return "  ?  "
    return f"{s:>4.2f}"


async def main():
    print("=" * 78)
    print("v15 SANITY TEST — 6 cases against the deepseek judge + reward server logic")
    print("=" * 78)

    if not os.environ.get("JUDGE_API_KEY"):
        print("ERROR: JUDGE_API_KEY not set. Source the secrets file first:")
        print('  source $WORK_ROOT/secrets/judge_api.env')
        sys.exit(2)

    total_pass = 0
    total_check = 0
    summaries = []

    for case in CASES:
        print()
        print("-" * 78)
        print(case["name"])
        print("-" * 78)
        try:
            result = await run_case(case)
        except Exception as e:
            print(f"  EXCEPTION: {e!r}")
            summaries.append((case["name"], 0, len(case["expected"])))
            total_check += len(case["expected"])
            continue

        eval_items = result.get("eval_items", [])
        cs = result.get("candidate_score", 0.0)
        fmt_factor = result.get("format_factor", 0.0)
        n_reuse_z = result.get("n_reuse_zeroed", 0)
        n_pairs = result.get("n_pairs", 0)
        total = result.get("score", 0.0)

        print(f"server: candidate_score={cs:.3f}, format_factor={fmt_factor}, "
              f"n_pairs={n_pairs}, n_reuse_zeroed={n_reuse_z}, total={total:.3f}")
        print()

        case_pass = 0
        for ev, exp in zip(eval_items, case["expected"]):
            label, hint, check = exp
            pm = ev.get("primary_match")
            sc = ev.get("score")
            crit = (ev.get("critique") or "")[:200].replace("\n", " ")
            ok = check(sc, pm)
            if ok:
                case_pass += 1
            verdict = "PASS" if ok else "FAIL"
            print(f"  [{verdict}] {label}")
            print(f"          expect: {hint}")
            print(f"          got:    score={fmt_score(sc)} pm={pm}")
            print(f"          critique: {crit}...")
        total_pass += case_pass
        total_check += len(case["expected"])
        summaries.append((case["name"], case_pass, len(case["expected"])))

    print()
    print("=" * 78)
    print("SUMMARY")
    print("=" * 78)
    for name, p, t in summaries:
        flag = "OK " if p == t else "** "
        print(f"  {flag}{name}: {p}/{t}")
    print()
    print(f"OVERALL: {total_pass}/{total_check} expectations met")
    if total_pass < total_check:
        print()
        print("If many FAILs, the judge is not following v15 prompt as designed.")
        print("Inspect the 'critique' text to see how the judge interpreted each case.")


if __name__ == "__main__":
    asyncio.run(main())