scripts/eval_gonyai.py

# ================================================================
#  GomParam-v1 × Gonyai-TEO2 Evaluation Script
#  Kaggle / Colab ready. Mirrors colab_eval.py harness exactly.
#
#  Run this first in a separate cell:
#  !pip -q install -U bitsandbytes accelerate transformers sentencepiece safetensors huggingface_hub
#
#  Usage:
#    python eval_gonyai.py
#    or paste into a Kaggle notebook cell after the pip install.
# ================================================================

import json, os, gc, time, math, logging
from pathlib import Path
from typing import Optional

import torch
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM

logging.basicConfig(level=logging.INFO, format="%(asctime)s  %(levelname)s  %(message)s")
log = logging.getLogger("gomparam_gonyai")

# ── CONFIG ─────────────────────────────────────────────────────
HF_TOKEN        = os.getenv("HF_TOKEN", "")          # set in Kaggle Secrets
GONYAI_REPO     = "omdeep22/gonyai-teo2"             # HF model repo
DATASET_REPO    = "omdeep22/GomParam-v1"             # benchmark repo
LOAD_IN_8BIT    = False                              # 251M fits in FP16 easily

IS_KAGGLE   = Path("/kaggle/working").exists()
BASE_DIR    = Path("/kaggle/working") if IS_KAGGLE else Path(".")
OUTPUT_DIR  = BASE_DIR / "gonyai_results"
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

# Route HF cache to working directory (avoids Kaggle disk OOM)
os.environ["HF_HOME"] = str(BASE_DIR / "hf_cache")

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE  = torch.float16 if DEVICE == "cuda" else torch.float32

print(f"Device : {DEVICE}")
if DEVICE == "cuda":
    print(f"GPU    : {torch.cuda.get_device_name(0)}")
    print(f"VRAM   : {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")

if HF_TOKEN:
    from huggingface_hub import login
    login(token=HF_TOKEN)
else:
    log.warning("No HF_TOKEN found. Set it in Kaggle Secrets if the repo is private.")


# ── 1. Load GomParam-v1 modules ────────────────────────────────
ALL_MODULES = [
    "cloze","code_switching","coherence","coreference","cross_scripting",
    "cultural_grounding","dialect","entailment","homograph_disambiguation",
    "idioms_proverbs","kinship","medical","mixed_general","morphology",
    "numerical_reasoning","para_qa","perplexity","pragmatics",
    "register_discrimination","sentiment","spatio_temporal",
]

print("\nLoading GomParam-v1 from HuggingFace...")
modules = {}
for m in ALL_MODULES:
    try:
        ds = load_dataset(DATASET_REPO, m, split="train", trust_remote_code=False,
                          token=HF_TOKEN or None)
        modules[m] = list(ds)
        print(f"  ✓ {m:35s} {len(modules[m]):>4} items")
    except Exception as e:
        print(f"  ✗ {m:35s} FAILED: {e}")

total = sum(len(v) for v in modules.values())
print(f"\nLoaded {len(modules)}/{len(ALL_MODULES)} modules — {total} total items\n")


# ── 2. Schema normaliser (identical to colab_eval.py) ──────────
def safe_str(*args, default=""):
    for v in args:
        if v is not None and str(v).strip():
            return str(v)
    return default

def normalise_mcq(item: dict, module: str) -> Optional[dict]:
    item = dict(item)
    candidates = item.get("candidates", [])
    correct    = item.get("correct", -1)
    if not candidates or correct == -1 or len(candidates) < 2:
        return None
    if all(not str(c).strip() for c in candidates):
        return None
    try:
        correct = int(correct)
    except (ValueError, TypeError):
        return None
    if correct < 0 or correct >= len(candidates):
        return None

    if module == "cloze":
        sentence = safe_str(item.get("sentence"), item.get("context"))
        if "___" in sentence:
            prefix = sentence.split("___")[0]
            suffix = sentence.split("___")[1] if len(sentence.split("___")) > 1 else ""
            cands  = [str(c).strip() + suffix for c in candidates]
        else:
            prefix = sentence
            cands  = [" " + str(c) for c in candidates]
        return {"prefix": prefix, "candidates": cands, "correct": correct}

    if module == "morphology":
        stem = safe_str(item.get("context"), item.get("sentence"))
        return {"prefix": stem, "candidates": [" " + str(c) for c in candidates], "correct": correct}

    if module == "entailment":
        premise    = safe_str(item.get("premise"),    item.get("context"))
        hypothesis = safe_str(item.get("hypothesis"), item.get("question"))
        prefix     = "पूर्वनिश्चय: " + premise + "\nप्रस्ताव: " + hypothesis
        return {"prefix": prefix, "candidates": ["\n" + str(c) for c in candidates], "correct": correct}

    if module == "idioms_proverbs":
        konkani  = safe_str(item.get("konkani"), item.get("context"), item.get("sentence"))
        question = safe_str(item.get("question"), "ह्या म्हणींचो अर्थ कितें?")
        prefix   = konkani + "\n" + question
        return {"prefix": prefix, "candidates": ["\n" + str(c) for c in candidates], "correct": correct}

    if module == "spatio_temporal":
        ctx = safe_str(item.get("context"), item.get("question"), item.get("sentence"))
        return {"prefix": ctx, "candidates": [" " + str(c) for c in candidates], "correct": correct}

    # Generic handler
    ctx      = safe_str(item.get("context"), item.get("passage"), item.get("sentence"),
                        item.get("scenario"), item.get("romi"), item.get("sentence_a"))
    question = safe_str(item.get("question"))
    prefix   = ctx + ("\n" + question if question else "")
    return {"prefix": prefix, "candidates": ["\n" + str(c) for c in candidates], "correct": correct}


# ── 3. Conditional log-probability scorer ──────────────────────
@torch.no_grad()
def sequence_log_prob(model, tokenizer, prefix: str, candidate: str, max_len: int = 512) -> float:
    """
    Conditional log P(candidate | prefix) per token.
    Prompt tokens are masked with -100 so only candidate loss is measured.
    Identical implementation to colab_eval.py — ensures directly comparable scores.
    """
    enc_full   = tokenizer(prefix + candidate, return_tensors="pt",
                           truncation=True, max_length=max_len).to(DEVICE)
    enc_prefix = tokenizer(prefix,             return_tensors="pt",
                           truncation=True, max_length=max_len)
    ids        = enc_full["input_ids"]
    prefix_len = enc_prefix["input_ids"].shape[1]
    if ids.shape[1] == 0 or prefix_len >= ids.shape[1]:
        return -1e9
    labels = ids.clone()
    labels[0, :prefix_len] = -100  # mask prompt tokens
    with torch.amp.autocast("cuda", enabled=(DEVICE == "cuda")):
        out = model(ids, labels=labels)
    return -float(out.loss.item())


def score_module(model, tokenizer, items, module: str) -> dict:
    correct, total = 0, 0
    diff_stats = {}
    for raw in items:
        raw = dict(raw)
        difficulty = str(raw.get("difficulty", "unknown")).lower().strip()
        norm = normalise_mcq(raw, module)
        if norm is None:
            continue
        lps  = [sequence_log_prob(model, tokenizer, norm["prefix"], c)
                for c in norm["candidates"]]
        pred = int(np.argmax(lps))
        gold = norm["correct"]
        hit  = int(pred == gold)
        correct += hit
        total   += 1
        diff_stats.setdefault(difficulty, {"correct": 0, "total": 0})
        diff_stats[difficulty]["total"]   += 1
        diff_stats[difficulty]["correct"] += hit

    diff_accuracy = {d: {"accuracy": s["correct"]/s["total"] if s["total"] else 0,
                         "correct": s["correct"], "total": s["total"]}
                     for d, s in diff_stats.items()}
    return {
        "accuracy": correct / total if total else 0.0,
        "correct":  correct,
        "total":    total,
        "by_difficulty": diff_accuracy,
    }


# ── 4. Module weights (same as colab_eval.py) ──────────────────
MODULE_WEIGHTS = {
    "morphology":0.15, "cloze":0.12, "para_qa":0.10, "idioms_proverbs":0.08,
    "pragmatics":0.08, "cultural_grounding":0.07, "homograph_disambiguation":0.07,
    "entailment":0.06, "coreference":0.06, "register_discrimination":0.05,
    "sentiment":0.04, "spatio_temporal":0.04, "kinship":0.04,
    "numerical_reasoning":0.03, "medical":0.03, "coherence":0.03,
    "cross_scripting":0.02, "code_switching":0.02, "dialect":0.02, "perplexity":0.02,
}
_total_w = sum(MODULE_WEIGHTS.values())
MODULE_WEIGHTS = {k: v / _total_w for k, v in MODULE_WEIGHTS.items()}
MCQ_MODULES = [m for m in MODULE_WEIGHTS if m in modules]


# ── 5. Load Gonyai-TEO2 ────────────────────────────────────────
print(f"Loading {GONYAI_REPO} ...")
tokenizer = AutoTokenizer.from_pretrained(GONYAI_REPO, use_fast=True,
                                          token=HF_TOKEN or None)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    GONYAI_REPO,
    torch_dtype=DTYPE,
    device_map={"": 0} if DEVICE == "cuda" else None,  # pin to cuda:0 — avoids cross-GPU split on Kaggle dual-T4
    trust_remote_code=True,
    token=HF_TOKEN or None,
)
if DEVICE != "cuda":
    model = model.to(DEVICE)
model.eval()

n_params = sum(p.numel() for p in model.parameters()) / 1e6
print(f"Model loaded: {n_params:.0f}M parameters\n")


# ── 6. Evaluate ────────────────────────────────────────────────
results = {}
for mod in MCQ_MODULES:
    t0 = time.time()
    sc = score_module(model, tokenizer, modules[mod], mod)
    dt = time.time() - t0

    diff_parts = []
    for d in ["basic", "intermediate", "advanced"]:
        if d in sc["by_difficulty"]:
            ds = sc["by_difficulty"][d]
            diff_parts.append(f"{d[:3]}:{ds['accuracy']*100:.0f}%")
    diff_str = "  [" + " | ".join(diff_parts) + "]" if diff_parts else ""

    log.info(f"  {mod:35s}  {sc['accuracy']*100:5.1f}%  "
             f"({sc['correct']:3d}/{sc['total']:3d})  {dt:.0f}s{diff_str}")
    results[mod] = sc


# ── 7. Composite score ─────────────────────────────────────────
def composite(res):
    ws, wt = 0.0, 0.0
    for mod, w in MODULE_WEIGHTS.items():
        if mod in res and "accuracy" in res[mod]:
            ws += res[mod]["accuracy"] * w
            wt += w
    return ws / wt if wt else 0.0

comp = composite(results)
print(f"\n{'='*60}")
print(f"Gonyai-TEO2 Composite Accuracy: {comp*100:.2f}%")
print(f"Random Baseline:                25.00%")
print(f"{'='*60}\n")

# Module table
print(f"{'Module':<35} {'Acc':>7}  {'Correct':>7}  {'Total':>7}")
print("-" * 60)
for mod in sorted(results):
    sc = results[mod]
    print(f"{mod:<35} {sc['accuracy']*100:>6.1f}%  {sc['correct']:>7}  {sc['total']:>7}")


# ── 8. Save results ────────────────────────────────────────────
summary = {
    "model":             GONYAI_REPO,
    "composite_accuracy": round(comp, 4),
    "random_baseline":    0.25,
    "per_module":        {m: {"accuracy": round(results[m]["accuracy"], 4),
                              "correct":  results[m]["correct"],
                              "total":    results[m]["total"],
                              "by_difficulty": results[m]["by_difficulty"]}
                         for m in results},
}
summary_path = OUTPUT_DIR / "gonyai_summary.json"
with open(summary_path, "w") as f:
    json.dump(summary, f, indent=2)
print(f"\nSummary saved → {summary_path}")


# ── 9. Heatmap plot ────────────────────────────────────────────
try:
    heat_data = {m: results[m]["accuracy"] * 100 for m in MCQ_MODULES}
    df_heat   = pd.DataFrame([heat_data], index=["Gonyai-TEO2"])
    df_heat   = df_heat.rename(columns={m: m.replace("_", "\n").title() for m in MCQ_MODULES})
    fig, ax   = plt.subplots(figsize=(18, 3))
    sns.heatmap(df_heat, annot=True, fmt=".0f", cmap="YlGnBu",
                vmin=0, vmax=100, linewidths=0.5, ax=ax,
                cbar_kws={"label": "Accuracy (%)"})
    ax.set_title("Gonyai-TEO2 — GomParam-v1 Per-Module Accuracy", fontsize=13, fontweight="bold")
    plt.xticks(rotation=45, ha="right", fontsize=8)
    plt.tight_layout()
    heatmap_path = OUTPUT_DIR / "gonyai_heatmap.png"
    plt.savefig(heatmap_path, dpi=200, bbox_inches="tight")
    plt.show()
    print(f"Heatmap saved → {heatmap_path}")
except Exception as e:
    log.warning(f"Heatmap plot failed: {e}")

print(f"\n✅ Done. Results in {OUTPUT_DIR}")