Upload runpod_benchmark.ipynb with huggingface_hub

runpod_benchmark.ipynb

@@ -0,0 +1,135 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "0p1yxckx3sxq",
+   "source": "# TAMIL-MORPH Benchmark — RunPod\n**Paper:** \"A Thousand Language Problem: Morphological Understanding in Linguistic AI\"\n\nBenchmarks 2 Tamil LLMs on 1,030 morphological test cases across 9 categories.\n- **Tamil-ai/tamil-qwen25-7b-instruct** (ours)\n- **abhinand/tamil-llama-13b-instruct-v0.1** (top Tamil open-source)",
+   "metadata": {}
+  },
+  {
+   "cell_type": "markdown",
+   "id": "4h2hutdkfg2",
+   "source": "## Step 1: Install Dependencies",
+   "metadata": {}
+  },
+  {
+   "cell_type": "code",
+   "id": "nbqryxy91u",
+   "source": "!pip install -q transformers accelerate bitsandbytes huggingface_hub sentencepiece protobuf",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7npo72pr67m",
+   "source": "## Step 2: Download Benchmark Data from HuggingFace",
+   "metadata": {}
+  },
+  {
+   "cell_type": "code",
+   "id": "uss67qa7bco",
+   "source": "from huggingface_hub import hf_hub_download\n\nbenchmark_path = hf_hub_download(\n    repo_id=\"Tamil-ai/tamil-morphological-benchmark\",\n    filename=\"Benchmarkdata.md\",\n    repo_type=\"dataset\",\n)\nprint(f\"Downloaded to: {benchmark_path}\")",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "id": "lnc0fhwtpcd",
+   "source": "## Step 3: Parse Benchmark Data (1,030 test cases)",
+   "metadata": {}
+  },
+  {
+   "cell_type": "code",
+   "id": "k7djjy65rj",
+   "source": "import json, re, gc, unicodedata, time\nfrom pathlib import Path\nfrom datetime import datetime\n\n# --- Config ---\nSYSTEM_PROMPT = (\n    \"You are a Tamil linguistics expert. \"\n    \"Answer with ONLY the Tamil word or phrase requested. \"\n    \"Do not add explanations, translations, or extra text. \"\n    \"Just the Tamil form, nothing else.\"\n)\n\nCATEGORY_NAMES = {\n    \"case_suffixes\": \"Case Suffixes (வேற்றுமை)\",\n    \"plural_case\": \"Plural + Case (பன்மை)\",\n    \"verb_conjugation\": \"Verb Conjugation (வினைத்திரிபு)\",\n    \"sandhi\": \"Sandhi (புணர்ச்சி)\",\n    \"honorific\": \"Honorific (மரியாதை)\",\n    \"negation\": \"Negation (எதிர்மறை)\",\n    \"compound\": \"Compound Words (கூட்டுச்சொல்)\",\n    \"conditional\": \"Conditional/Causal (நிபந்தனை)\",\n    \"novel\": \"Novel Combinations (புதிய வடிவங்கள்)\",\n}\n\nCASE_NAMES_TAMIL = {\n    \"accusative\": \"இரண்டாம் வேற்றுமை (accusative / -ஐ)\",\n    \"dative\": \"நான்காம் வேற்றுமை (dative / -க்கு)\",\n    \"locative\": \"ஏழாம் வேற்றுமை (locative / -இல்)\",\n    \"ablative\": \"ablative (-இலிருந்து)\",\n    \"genitive\": \"ஆறாம் வேற்றுமை (genitive / -இன்)\",\n    \"sociative\": \"மூன்றாம் வேற்றுமை (sociative / -ஓடு)\",\n}\n\nPERSON_TAMIL = {\n    \"naan_present\": (\"நான்\", \"present tense / நிகழ்காலம்\"),\n    \"naan_past\": (\"நான்\", \"past tense / இறந்தகாலம்\"),\n    \"naan_future\": (\"நான்\", \"future tense / எதிர்காலம்\"),\n    \"avan_present\": (\"அவன்\", \"present tense / நிகழ்காலம்\"),\n    \"aval_present\": (\"அவள்\", \"present tense / நிகழ்காலம்\"),\n    \"avargal_present\": (\"அவர்கள்\", \"present tense / நிகழ்காலம்\"),\n    \"naangal_present\": (\"நாங்கள்\", \"present tense / நிகழ்காலம்\"),\n}\n\nPLURAL_FORMS = {\n    \"plural\": \"plural / பன்மை\",\n    \"plural_locative\": \"plural + locative / பன்மை + -இல்\",\n    \"plural_ablative\": \"plural + ablative / பன்மை + -இலிருந்து\",\n    \"plural_dative\": \"plural + dative / பன்மை + -க்கு\",\n}\n\nNEGATION_FORMS = {\n    \"present_negative\": \"present negative / நிகழ்கால எதிர்மறை\",\n    \"past_negative\": \"past negative / இறந்தகால எதிர்மறை\",\n    \"future_negative\": \"future negative / எதிர்கால எதிர்மறை\",\n}\n\nCONDITIONAL_FORMS = {\n    \"conditional\": \"conditional / நிபந்தனை (-ஆல்)\",\n    \"causal\": \"causal / காரண (-ததால்)\",\n}\n\nHONORIFIC_LEVELS = {\n    \"informal\": \"informal / முறைசாரா (நீ)\",\n    \"formal\": \"formal / மரியாதை (நீங்கள்)\",\n    \"high_honorific\": \"high honorific / உயர் மரியாதை (literary)\",\n}\n\n# --- Parser ---\nSECTION_MARKERS = [\n    (r\"\\*\\*## PROMPT 1.*Case Suffixes\", \"case_suffixes\"),\n    (r\"## PROMPT 2.*Plural\", \"plural_case\"),\n    (r\"## PROMPT 3.*Verb Conjugation\", \"verb_conjugation\"),\n    (r\"\\*\\*prompt 4\\*\\*\", \"sandhi\"),\n    (r\"\\*\\*prompt 5\\*\\*\", \"honorific\"),\n    (r\"\\*\\*prompt 6\\*\\*\", \"negation\"),\n    (r\"\\*\\*prompt 7\\*\\*\", \"compound\"),\n    (r\"\\*\\*prompt 8\\*\\*\", \"conditional\"),\n    (r\"\\*\\*prompt\\*\\*\", \"novel\"),\n]\n\ndef extract_json_blocks(text):\n    blocks, depth, start = [], 0, None\n    for i, ch in enumerate(text):\n        if ch == \"[\" and depth == 0: start, depth = i, 1\n        elif ch == \"[\": depth += 1\n        elif ch == \"]\":\n            depth -= 1\n            if depth == 0 and start is not None:\n                blocks.append(text[start:i+1]); start = None\n    return blocks\n\ndef parse_benchmark_data(filepath):\n    raw = Path(filepath).read_text(encoding=\"utf-8\")\n    lines = raw.split(\"\\n\")\n    sections, current_key, current_lines = {}, None, []\n    for line in lines:\n        matched = False\n        for pattern, key in SECTION_MARKERS:\n            if re.search(pattern, line, re.IGNORECASE):\n                if current_key: sections[current_key] = \"\\n\".join(current_lines)\n                current_key, current_lines, matched = key, [], True; break\n        if not matched and current_key: current_lines.append(line)\n    if current_key: sections[current_key] = \"\\n\".join(current_lines)\n\n    parsed = {}\n    for key, content in sections.items():\n        blocks = extract_json_blocks(content)\n        if blocks:\n            try: parsed[key] = json.loads(blocks[0])\n            except json.JSONDecodeError as e: print(f\"  WARNING: {key}: {e}\"); parsed[key] = []\n    return parsed\n\nprint(\"Parsing benchmark data...\")\ndata = parse_benchmark_data(benchmark_path)\nfor k, v in data.items():\n    print(f\"  {CATEGORY_NAMES.get(k, k)}: {len(v)} items\")",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "id": "od2il1scq5c",
+   "source": "## Step 4: Generate Test Cases",
+   "metadata": {}
+  },
+  {
+   "cell_type": "code",
+   "id": "u2e7qm5s5do",
+   "source": "def generate_test_cases(data):\n    cases = []\n\n    for item in data.get(\"case_suffixes\", []):\n        root, meaning = item[\"root\"], item[\"root_meaning\"]\n        for case_key, form_data in item[\"forms\"].items():\n            cases.append({\n                \"category\": \"case_suffixes\", \"sub_id\": f\"{root}_{case_key}\",\n                \"prompt\": f\"What is the {case_key} form of the Tamil word '{root}' ({meaning})?\\nTamil grammar term: {CASE_NAMES_TAMIL.get(case_key, case_key)}\\nExample: 'வீடு' (house) accusative → வீட்டை\\nAnswer with ONLY the Tamil word.\",\n                \"expected\": form_data[\"tamil\"], \"root\": root,\n            })\n\n    for item in data.get(\"plural_case\", []):\n        root, meaning = item[\"root\"], item[\"root_meaning\"]\n        for form_key, form_data in item[\"forms\"].items():\n            cases.append({\n                \"category\": \"plural_case\", \"sub_id\": f\"{root}_{form_key}\",\n                \"prompt\": f\"What is the {PLURAL_FORMS.get(form_key, form_key)} form of the Tamil word '{root}' ({meaning})?\\nExample: 'வீடு' plural → வீடுகள், plural+locative → வீடுகளில்\\nAnswer with ONLY the Tamil word.\",\n                \"expected\": form_data[\"tamil\"], \"root\": root,\n            })\n\n    for item in data.get(\"verb_conjugation\", []):\n        root, meaning = item[\"root\"], item[\"root_meaning\"]\n        for form_key, form_data in item[\"forms\"].items():\n            if form_key in PERSON_TAMIL:\n                person, tense_desc = PERSON_TAMIL[form_key]\n                prompt = f\"Conjugate the Tamil verb '{root}' ({meaning}) for {person} in {tense_desc}.\\nExample: 'படி' (read), நான், present → படிக்கிறேன்\\nAnswer with the full form (e.g., படிக்கிறேன், not just the verb).\"\n            else:\n                prompt = f\"Conjugate the Tamil verb '{root}' ({meaning}) in the {form_key} form.\\nAnswer with ONLY the Tamil form.\"\n            cases.append({\n                \"category\": \"verb_conjugation\", \"sub_id\": f\"{root}_{form_key}\",\n                \"prompt\": prompt, \"expected\": form_data[\"tamil\"], \"root\": root,\n            })\n\n    for item in data.get(\"sandhi\", []):\n        w1, m1 = item.get(\"word1\", \"\"), item.get(\"word1_meaning\", \"\")\n        w2, m2 = item.get(\"word2\", \"\"), item.get(\"word2_meaning\", \"\")\n        expected = item.get(\"combined\", \"\")\n        if not expected or not w1: continue\n        cases.append({\n            \"category\": \"sandhi\", \"sub_id\": f\"{w1}+{w2}\",\n            \"prompt\": f\"When the Tamil words '{w1}' ({m1}) and '{w2}' ({m2}) combine according to Tamil sandhi (புணர்ச்சி) rules, what is the combined form?\\nExample: 'மலை' + 'அழகு' → மலையழகு\\nAnswer with ONLY the combined Tamil word.\",\n            \"expected\": expected, \"root\": f\"{w1}+{w2}\",\n        })\n\n    for item in data.get(\"honorific\", []):\n        action = item.get(\"action\", \"\")\n        for level_key, form_data in item.get(\"forms\", {}).items():\n            cases.append({\n                \"category\": \"honorific\", \"sub_id\": f\"{action}_{level_key}\",\n                \"prompt\": f\"What is the {HONORIFIC_LEVELS.get(level_key, level_key)} Tamil form for the action '{action}'?\\nExample: 'come' informal → வா, formal → வாருங்கள்\\nAnswer with ONLY the Tamil word or short phrase.\",\n                \"expected\": form_data[\"tamil\"], \"root\": action,\n            })\n\n    for item in data.get(\"negation\", []):\n        root, meaning = item[\"root\"], item[\"root_meaning\"]\n        for form_key, form_data in item[\"forms\"].items():\n            cases.append({\n                \"category\": \"negation\", \"sub_id\": f\"{root}_{form_key}\",\n                \"prompt\": f\"What is the {NEGATION_FORMS.get(form_key, form_key)} form of the Tamil verb '{root}' ({meaning}) for நான் (I)?\\nExample: 'படி' (read), present negative → படிக்கவில்லை\\nAnswer with the full Tamil phrase.\",\n                \"expected\": form_data[\"tamil\"], \"root\": root,\n            })\n\n    for item in data.get(\"compound\", []):\n        w1, m1 = item.get(\"word1\", \"\"), item.get(\"word1_meaning\", \"\")\n        w2, m2 = item.get(\"word2\", \"\"), item.get(\"word2_meaning\", \"\")\n        expected = item.get(\"compound\", \"\")\n        if not expected or not w1: continue\n        cases.append({\n            \"category\": \"compound\", \"sub_id\": f\"{w1}+{w2}\",\n            \"prompt\": f\"What is the Tamil compound word formed by combining '{w1}' ({m1}) and '{w2}' ({m2})?\\nExample: 'தலை' (head) + 'வலி' (pain) → தலைவலி (headache)\\nAnswer with ONLY the compound Tamil word.\",\n            \"expected\": expected, \"root\": f\"{w1}+{w2}\",\n        })\n\n    for item in data.get(\"conditional\", []):\n        root, meaning = item[\"root\"], item[\"root_meaning\"]\n        for form_key, form_data in item[\"forms\"].items():\n            cases.append({\n                \"category\": \"conditional\", \"sub_id\": f\"{root}_{form_key}\",\n                \"prompt\": f\"What is the {CONDITIONAL_FORMS.get(form_key, form_key)} form of the Tamil verb '{root}' ({meaning})?\\nExample: 'படி' (read), conditional → படித்தால், causal → படித்ததால்\\nAnswer with ONLY the Tamil word.\",\n                \"expected\": form_data[\"tamil\"], \"root\": root,\n            })\n\n    for item in data.get(\"novel\", []):\n        cases.append({\n            \"category\": \"novel\", \"sub_id\": f\"novel_{item.get('id','')}_{item.get('category','novel')}\",\n            \"prompt\": f\"Combine the following Tamil morphemes into a single valid Tamil word/form:\\nMorphemes: {item['breakdown']}\\nIntended meaning: {item['meaning']}\\nExample: 'பல்கலைக்கழகம் + கள் + இல் + இருந்து + கூட' → பல்கலைக்கழகங்களிலிருந்துகூட\\nAnswer with ONLY the combined Tamil form.\",\n            \"expected\": item[\"form\"], \"root\": item[\"breakdown\"],\n        })\n\n    return cases\n\ntest_cases = generate_test_cases(data)\nprint(f\"Total test cases: {len(test_cases)}\")\nfor cat_key in CATEGORY_NAMES:\n    count = sum(1 for tc in test_cases if tc[\"category\"] == cat_key)\n    if count > 0:\n        print(f\"  {CATEGORY_NAMES[cat_key]}: {count}\")",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f9y1w141jhi",
+   "source": "## Step 5: Scoring & Inference Functions",
+   "metadata": {}
+  },
+  {
+   "cell_type": "code",
+   "id": "gixhah0g5h",
+   "source": "import torch\nfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\n\ndef normalize_tamil(text):\n    text = unicodedata.normalize(\"NFC\", text).strip()\n    text = re.sub(r'[\"\"\"\\'`\\.\\,\\!\\?\\;\\:\\(\\)\\[\\]\\{\\}]', '', text)\n    return text.strip()\n\ndef extract_tamil_answer(response):\n    response = response.strip()\n    tamil_pattern = re.compile(r'[\\u0B80-\\u0BFF][\\u0B80-\\u0BFF\\s]*[\\u0B80-\\u0BFF]')\n    single_tamil = re.compile(r'[\\u0B80-\\u0BFF]+')\n    for line in response.split(\"\\n\"):\n        line = line.strip()\n        if not line: continue\n        if any(line.startswith(p) for p in [\"Note:\", \"Explanation:\", \"The \", \"This \", \"Here\"]): continue\n        match = tamil_pattern.search(line)\n        if match: return normalize_tamil(match.group())\n        match = single_tamil.search(line)\n        if match: return normalize_tamil(match.group())\n    match = tamil_pattern.search(response)\n    if match: return normalize_tamil(match.group())\n    match = single_tamil.search(response)\n    if match: return normalize_tamil(match.group())\n    return normalize_tamil(response)\n\ndef score_response(expected, predicted):\n    exp_norm, pred_norm = normalize_tamil(expected), normalize_tamil(predicted)\n    if not pred_norm: return 0.0\n    if exp_norm == pred_norm: return 1.0\n    if exp_norm in pred_norm: return 1.0\n    if pred_norm in exp_norm and len(pred_norm) > 2: return 0.5\n    return 0.0\n\ndef load_model(model_id):\n    print(f\"\\nLoading {model_id}...\")\n    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)\n    if tokenizer.pad_token is None:\n        tokenizer.pad_token = tokenizer.eos_token\n    model = AutoModelForCausalLM.from_pretrained(\n        model_id, trust_remote_code=True, device_map=\"auto\",\n        torch_dtype=torch.float16,\n        quantization_config=BitsAndBytesConfig(\n            load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16, bnb_4bit_quant_type=\"nf4\",\n        ),\n    )\n    model.eval()\n    print(f\"  Loaded on {model.device}\")\n    return model, tokenizer\n\ndef generate_response(model, tokenizer, prompt):\n    messages = [\n        {\"role\": \"system\", \"content\": SYSTEM_PROMPT},\n        {\"role\": \"user\", \"content\": prompt},\n    ]\n    try:\n        text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n    except Exception:\n        text = f\"{SYSTEM_PROMPT}\\n\\nUser: {prompt}\\nAssistant:\"\n    inputs = tokenizer(text, return_tensors=\"pt\").to(model.device)\n    with torch.no_grad():\n        outputs = model.generate(\n            **inputs, max_new_tokens=64, temperature=0.1, top_p=0.9,\n            do_sample=True, repetition_penalty=1.1, pad_token_id=tokenizer.pad_token_id,\n        )\n    return tokenizer.decode(outputs[0][inputs[\"input_ids\"].shape[1]:], skip_special_tokens=True).strip()\n\ndef free_model(model, tokenizer):\n    del model, tokenizer\n    gc.collect()\n    torch.cuda.empty_cache()\n    print(\"  GPU memory freed.\")\n\ndef run_evaluation(model, tokenizer, test_cases):\n    results = []\n    for i, tc in enumerate(test_cases):\n        raw = generate_response(model, tokenizer, tc[\"prompt\"])\n        predicted = extract_tamil_answer(raw)\n        sc = score_response(tc[\"expected\"], predicted)\n        results.append({\n            \"category\": tc[\"category\"], \"sub_id\": tc[\"sub_id\"],\n            \"expected\": tc[\"expected\"], \"predicted\": predicted,\n            \"raw_response\": raw[:200], \"score\": sc,\n        })\n        if (i + 1) % 50 == 0 or i == 0:\n            cat_so_far = [r for r in results if r[\"category\"] == tc[\"category\"]]\n            cat_acc = sum(r[\"score\"] for r in cat_so_far) / len(cat_so_far) if cat_so_far else 0\n            print(f\"  [{i+1}/{len(test_cases)}] {tc['category']}: {cat_acc:.1%}\")\n    return results\n\ndef generate_report(results, model_name):\n    categories = {}\n    for r in results:\n        cat = r[\"category\"]\n        if cat not in categories:\n            categories[cat] = {\"total\": 0, \"exact\": 0, \"partial\": 0, \"wrong\": 0, \"score_sum\": 0}\n        categories[cat][\"total\"] += 1\n        categories[cat][\"score_sum\"] += r[\"score\"]\n        if r[\"score\"] == 1.0: categories[cat][\"exact\"] += 1\n        elif r[\"score\"] == 0.5: categories[cat][\"partial\"] += 1\n        else: categories[cat][\"wrong\"] += 1\n\n    print(f\"\\n{'='*80}\")\n    print(f\"RESULTS: {model_name}\")\n    print(f\"{'='*80}\")\n    print(f\"{'Category':<40} {'Total':>6} {'Exact':>6} {'Partial':>8} {'Wrong':>6} {'Acc':>9}\")\n    print(f\"{'-'*80}\")\n    overall_score, overall_total, cat_scores = 0, 0, {}\n    for cat_key in CATEGORY_NAMES:\n        if cat_key in categories:\n            c = categories[cat_key]\n            acc = c[\"score_sum\"] / c[\"total\"] if c[\"total\"] > 0 else 0\n            print(f\"{CATEGORY_NAMES[cat_key]:<40} {c['total']:>6} {c['exact']:>6} {c['partial']:>8} {c['wrong']:>6} {acc:>8.1%}\")\n            overall_score += c[\"score_sum\"]; overall_total += c[\"total\"]\n            cat_scores[cat_key] = round(acc * 100, 1)\n    overall_acc = overall_score / overall_total if overall_total > 0 else 0\n    print(f\"{'-'*80}\")\n    print(f\"{'OVERALL':<40} {overall_total:>6} {'':>6} {'':>8} {'':>6} {overall_acc:>8.1%}\")\n    print(f\"{'='*80}\")\n    failures = [r for r in results if r[\"score\"] == 0.0]\n    if failures:\n        print(f\"\\nSample failures (first 10):\")\n        for f in failures[:10]:\n            print(f\"  [{f['category']}] expected='{f['expected']}' got='{f['predicted']}'\")\n    return {\"model\": model_name, \"overall_accuracy\": round(overall_acc * 100, 1),\n            \"total_cases\": overall_total, \"category_scores\": cat_scores}\n\nprint(\"Functions ready.\")",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "id": "1upjcl4v6fm",
+   "source": "## Step 6: Run Tamil-Qwen 7B (Our Model)\n~15-20 min on RTX 5090",
+   "metadata": {}
+  },
+  {
+   "cell_type": "code",
+   "id": "hidsobegji8",
+   "source": "model_1, tok_1 = load_model(\"Tamil-ai/tamil-qwen25-7b-instruct\")\nresults_qwen = run_evaluation(model_1, tok_1, test_cases)\nreport_qwen = generate_report(results_qwen, \"Tamil-ai/tamil-qwen25-7b-instruct\")\n\n# Save results\nwith open(\"tamil_qwen_7b_results.json\", \"w\", encoding=\"utf-8\") as f:\n    json.dump(results_qwen, f, ensure_ascii=False, indent=2)\nprint(\"\\nSaved: tamil_qwen_7b_results.json\")\n\nfree_model(model_1, tok_1)",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "id": "fj37ldrjm89",
+   "source": "## Step 7: Run Tamil-Llama 13B (Top Tamil Open-Source)\n~20-25 min on RTX 5090",
+   "metadata": {}
+  },
+  {
+   "cell_type": "code",
+   "id": "5w3ap482t8o",
+   "source": "model_2, tok_2 = load_model(\"abhinand/tamil-llama-13b-instruct-v0.1\")\nresults_llama = run_evaluation(model_2, tok_2, test_cases)\nreport_llama = generate_report(results_llama, \"abhinand/tamil-llama-13b-instruct-v0.1\")\n\n# Save results\nwith open(\"tamil_llama_13b_results.json\", \"w\", encoding=\"utf-8\") as f:\n    json.dump(results_llama, f, ensure_ascii=False, indent=2)\nprint(\"\\nSaved: tamil_llama_13b_results.json\")\n\nfree_model(model_2, tok_2)",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "id": "lld5nwdhbt",
+   "source": "## Step 8: Comparison Table (All Models Including GPT-4o-mini)",
+   "metadata": {}
+  },
+  {
+   "cell_type": "code",
+   "id": "1omd5o1ow5n",
+   "source": "# GPT-4o-mini baseline (already benchmarked locally)\nreport_gpt = {\n    \"model\": \"gpt-4o-mini\",\n    \"overall_accuracy\": 50.7,\n    \"total_cases\": 1030,\n    \"category_scores\": {\n        \"case_suffixes\": 82.1, \"plural_case\": 38.1, \"verb_conjugation\": 45.0,\n        \"sandhi\": 42.0, \"honorific\": 30.0, \"negation\": 28.3,\n        \"compound\": 62.0, \"conditional\": 68.3, \"novel\": 30.0,\n    }\n}\n\nall_reports = [report_qwen, report_llama, report_gpt]\n\n# Print comparison\nprint(f\"\\n{'='*100}\")\nprint(\"TAMIL-MORPH BENCHMARK COMPARISON (1,030 test cases)\")\nprint(f\"{'='*100}\")\nheader = f\"{'Category':<35}\"\nfor r in all_reports:\n    short = r['model'].split('/')[-1][:22]\n    header += f\" {short:>22}\"\nprint(header)\nprint(\"-\" * 100)\nfor cat_key, cat_name in CATEGORY_NAMES.items():\n    row = f\"{cat_name[:35]:<35}\"\n    for report in all_reports:\n        score = report[\"category_scores\"].get(cat_key, \"-\")\n        if isinstance(score, (int, float)):\n            row += f\" {score:>21.1f}%\"\n        else:\n            row += f\" {'N/A':>22}\"\n    print(row)\nprint(\"-\" * 100)\nrow = f\"{'OVERALL':<35}\"\nfor report in all_reports:\n    row += f\" {report['overall_accuracy']:>21.1f}%\"\nprint(row)\nprint(\"=\" * 100)\n\n# Save combined\nwith open(\"benchmark_comparison.json\", \"w\", encoding=\"utf-8\") as f:\n    json.dump(all_reports, f, ensure_ascii=False, indent=2)\nprint(\"\\nSaved: benchmark_comparison.json\")",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "name": "python",
+   "version": "3.10.0"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}