File size: 16,877 Bytes
fe75945
61512c1
fe75945
 
 
 
 
 
 
 
 
f1566df
fe75945
 
f1566df
 
fe75945
51db0e0
 
db891fa
51db0e0
f1566df
7234c27
689986a
fe75945
51db0e0
 
 
fe75945
51db0e0
fe75945
 
f1566df
14f5c79
fe75945
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
51db0e0
fe75945
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
51db0e0
fe75945
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
542577d
fe75945
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
#Ψ¨Ψ³Ω… Ψ§Ω„Ω„Ω‡ Ψ§Ω„Ψ±Ψ­Ω…Ω† Ψ§Ω„Ψ±Ψ­ΩŠΩ…
from fastmcp import FastMCP
import asyncio
import json
import os
import time
from langchain_core.prompts import ChatPromptTemplate
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_groq import ChatGroq
from dotenv import load_dotenv
import nltk
import sys
from modal_tool import train_with_modal, app, upload_to_hf_from_volume, evaluate_model

mcp = FastMCP(name="FistalMCP")


groq = os.getenv("GROQ_API_KEY")
hf = os.getenv("HF_TOKEN")

if not groq:
    print("GROQ_API_KEY missing!", file=sys.stderr)
if not hf:
    print("HF Token not valid", file=sys.stderr)

gk1 = os.environ.get("GOOGLE_API_KEY_1")
gk2 = os.environ.get("GOOGLE_API_KEY_2")
gk3 = os.environ.get("GOOGLE_API_KEY_3")

GOOGLE_API_KEYS = [k for k in [gk1, gk2, gk3] if k]

if not GOOGLE_API_KEYS:
    print("No Google API keys found!", file=sys.stderr)



async def genBatch(topic: str, samples_per_batch: int, batch_num: int, api_key: str, task_type: str) -> list:
    """Generate one batch of samples using a single API key"""
    
    if not api_key or api_key == "YOUR_API_KEY":
        return []
    
    llm = ChatGoogleGenerativeAI(
        model="gemini-2.5-flash",
        temperature=0.7,
        google_api_key=api_key
    )

    prompt_template = """
You are an expert dataset generator.
Generate authentic, high-quality data on the topic: {topic} for task type: {task_type} using your knowledge.
Generate exactly {num} concise, varied, and high-quality samples.
Return a JSON list of objects, each with keys: instruction, input, and output.
Do not add extra texts, markdown, or code fences.
RESPONSE:
"""

    promptJSON = ChatPromptTemplate.from_template(prompt_template)
    chain = promptJSON | llm

    try:
        user_input = {
            "topic": topic,
            "num": samples_per_batch,
            "task_type": task_type
        }

        response = await asyncio.to_thread(chain.invoke, user_input)
        content = response.content.strip()

        if content.startswith("```json"):
            content = content[7:]
        if content.startswith("```"):
            content = content[3:]
        if content.endswith("```"):
            content = content[:-3]

        content = content.strip()
        data = json.loads(content)

        return data if isinstance(data, list) else [data]
    
    except json.JSONDecodeError as e:
        print(f"JSON decode error in batch {batch_num}: {e}")
        return []
    except Exception as e:
        print(f"Error in batch {batch_num}: {e}")
        return []


@mcp.tool()
async def generate_json_data(topic: str,  task_type: str, num_samples: int = 1000) -> str:
    """
    Generate a training dataset with instruction, input, and output fields.
    Uses parallel batching for efficiency. Can generate up to 2000 samples.
    
    Args:
        topic: The topic or theme for the dataset
        num_samples: Number of training examples to generate (recommended: 100-2000)
    
    Returns:
        JSON string with status, topic, total_samples, and data array
    """
    topic = str(topic).strip() if topic else ""
    task_type = str(task_type).strip() if task_type else "text-generation"
    
    try:
        num_samples = int(num_samples)
    except (ValueError, TypeError):
        num_samples = 100

    if not topic:
        return json.dumps({
            "status": "error",
            "message": "Topic cannot be empty"
        })
    if num_samples <= 0 or num_samples > 2000:
        num_samples = min(max(50, num_samples), 2000)

    
    valid_keys = [k for k in GOOGLE_API_KEYS if k and k.strip() and k != "YOUR_API_KEY"]
    if not valid_keys:
        return json.dumps({
            "status": "error",
            "message": "No valid Google API keys configured"
        })

    start_time = time.time()
    samples_per_batch = 50
    total_batches = (num_samples + samples_per_batch - 1) // samples_per_batch

    try:
        tasks = []

        for batch_num in range(total_batches):
            api_key = valid_keys[batch_num % len(valid_keys)]
            task = genBatch(
                topic=topic.strip(),
                samples_per_batch=samples_per_batch,
                batch_num=batch_num + 1,
                api_key=api_key,
                task_type=task_type.strip()
            )
            tasks.append(task)

        results = await asyncio.gather(*tasks, return_exceptions=True)

        all_samples = []
        for batch_result in results:
            if isinstance(batch_result, Exception):
                continue
            if isinstance(batch_result, list):
                all_samples.extend(batch_result)

        all_samples = all_samples[:num_samples]
        end_time = time.time()
        gen_time = end_time - start_time

        return json.dumps({
            "status": "success",
            "topic": topic,
            "task_type": task_type,
            "total_samples": len(all_samples),
            "requested_samples": num_samples,
            "total_batches": total_batches,
            "generation_time_seconds": round(gen_time, 1),
            "generation_time_minutes": round(gen_time / 60, 2),
            "samples_per_second": round(len(all_samples) / gen_time, 2) if gen_time > 0 else 0,
            "data": all_samples
        })
        
    except Exception as e:
        return json.dumps({
            "status": "error",
            "message": f"Error generating dataset: {str(e)}"
        })


@mcp.tool()
async def format_json(raw_data) -> str:  
    """
    Convert raw dataset to ChatML format for training
    
    Args:
        raw_data: List or JSON string of samples with instruction/input/output
    
    Returns:
        JSON string with status, num_samples, and formatted data
    """
    try:
        if isinstance(raw_data, list):
            data = raw_data
        elif isinstance(raw_data, str):
            parsed = json.loads(raw_data)
            if isinstance(parsed, dict) and "data" in parsed:
                data = parsed["data"]
            else:
                data = parsed
        elif isinstance(raw_data, dict) and "data" in raw_data:
            data = raw_data["data"]
        else:
            return json.dumps({
                "status": "error",
                "message": f"Unexpected input type: {type(raw_data).__name__}"
            })
        
        if not isinstance(data, list):
            return json.dumps({
                "status": "error",
                "message": "Data must be a list of samples"
            })

        # Convert to ChatML format
        converted = []
        for item in data:
            if not isinstance(item, dict):
                continue
            
            if 'instruction' not in item or 'output' not in item:
                continue

            user_msg = str(item['instruction'])
            if item.get('input'):
                user_msg += f"\n\n{item['input']}"
            
            converted.append({
                "messages": [
                    {"role": "system", "content": "You are a helpful assistant."},
                    {"role": "user", "content": user_msg},
                    {"role": "assistant", "content": str(item['output'])}
                ]
            })

        if not converted:
            return json.dumps({
                "status": "error",
                "message": "No valid samples to format"
            })

        return json.dumps({
            "status": "success",
            "num_samples": len(converted),
            "data": converted,
            "message": f"βœ… Formatted {len(converted)} samples"
        }, ensure_ascii=False)
        
    except Exception as e:
        import traceback
        return json.dumps({
            "status": "error",
            "message": f"Formatting failed: {str(e)}",
            "traceback": traceback.format_exc()
        })
        


@mcp.tool()
async def finetune_model(formatted_data, model_name: str, topic: str, task_type: str) -> str:  
    """
    Fine-tune model on Modal GPU
    
    Args:
        formatted_data: List or JSON string with formatted training samples
        model_name: Base model to fine-tune
    
    Returns:
        JSON string with status, repo_id, model_url
    """
    model_name = str(model_name).strip()
    
    models = [
                            "unsloth/Llama-3.2-1B-Instruct-bnb-4bit",
                            "unsloth/Phi-3-mini-4k-instruct",
                            "unsloth/Phi-3-medium-4k-instruct",
                            "unsloth/Llama-3.2-3B-Instruct-bnb-4bit",
                            "unsloth/Qwen2.5-3B-Instruct-bnb-4bit",
                            "unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit",
                            "unsloth/Qwen2.5-0.5B-Instruct-bnb-4bit",
                            "unsloth/Qwen2.5-Coder-3B-Instruct-bnb-4bit",
                            "unsloth/gemma-2-2b-it-bnb-4bit",
                            "unsloth/SmolLM2-1.7B-Instruct-bnb-4bit",
                            "unsloth/Phi-3.5-mini-instruct-bnb-4bit",
                            "unsloth/Granite-3.0-2b-instruct-bnb-4bit",
                            "unsloth/granite-4.0-h-1b-bnb-4bit"
    ]
    
    if model_name not in models:
        return json.dumps({
            "status": "error",
            "message": f"Model not supported. Choose from: {', '.join(models[:3])}..."
        })
    
    try:
        if isinstance(formatted_data, list):
            training_data = formatted_data
        elif isinstance(formatted_data, str):
            parsed = json.loads(formatted_data)
            if isinstance(parsed, dict) and "data" in parsed:
                training_data = parsed["data"]
            else:
                training_data = parsed
        elif isinstance(formatted_data, dict) and "data" in formatted_data:
            training_data = formatted_data["data"]
        else:
            return json.dumps({
                "status": "error",
                "message": f"Unexpected input type: {type(formatted_data).__name__}"
            })
        
        if not isinstance(training_data, list) or not training_data:
            return json.dumps({
                "status": "error",
                "message": "No training samples provided"
            })
        
        jsonl_content = "\n".join([json.dumps(s, ensure_ascii=False) for s in training_data])

        with app.run():
            result = train_with_modal.remote(jsonl_content, model_name)
        
        if result["status"] != "success":
            return json.dumps({
                "status": "error",
                "message": "Training failed"
            })
        
        repoTemp = """
Generate a short repository name for an unsloth finetuned model based on {topic} and {task_type}. 
Use '_' instead of spaces. Only return the name without quotations.
"""
        repoPrompt = ChatPromptTemplate.from_template(repoTemp)
        llm = ChatGroq(
        model="llama-3.1-8b-instant",
        temperature=0.4,
        api_key=groq
    )

        chain = repoPrompt | llm

        inp = {
            "topic": topic,
            "task_type": task_type
        }

        repoName = await asyncio.to_thread(chain.invoke, inp)
        repoName = repoName.content.strip()


        
        with app.run():
            hf_result = upload_to_hf_from_volume.remote(
                result["volume_path"], 
                result["timestamp"],
                repoName
            )

        return json.dumps({
            "status": "success",
            "repo_id": str(hf_result["repo_id"]),
            "model_url": str(hf_result["model_url"]),
            "model_path": str(hf_result["repo_id"]),
            "num_samples": len(training_data),
            "message": f"βœ… Model at {hf_result['model_url']}"
        })

    except Exception as e:
        import traceback
        return json.dumps({
            "status": "error",
            "message": f"Training failed: {str(e)}",
            "traceback": traceback.format_exc()
        })


@mcp.tool()
async def llm_as_judge(repo_id:str, topic: str, task_type: str) -> dict:
    """Use LLM to judge model quality based on topic and task type"""
    import evaluate
    eval_llm = ChatGroq(
        model="llama-3.1-8b-instant",
        temperature=0.2,
        api_key=groq
    )
    test_prompt_text = f"""Generate 3 test cases for evaluating a model fine-tuned strictly based on **{topic} for {task_type}**.
Return ONLY a JSON array with this exact format, no other text:
[{{"input": "test question 1", "expected_output": "expected answer 1"}}, {{"input": "test question 2", "expected_output": "expected answer 2"}}, {{"input": "test question 3", "expected_output": "expected answer 3"}}]"""
    try:
        text_responses = await eval_llm.ainvoke(test_prompt_text)
        response = text_responses.content.strip()
        response = response.replace("```json", "").replace("```", "").strip()
        import re
        match = re.search(r'\[.*\]', response, re.DOTALL)
        if match:
            response = match.group(0)
        
        test_cases = json.loads(response)[:3]

        test_inputs = [case['input'] for case in test_cases]

        with app.run():
            ft_output = evaluate_model.remote(repo_id, test_inputs)

        outputs = []
        for i, case in enumerate(test_cases):
            outputs.append(
                {
                    "input": case['input'],
                    "expected_output": case['expected_output'],
                    "model_output": ft_output[i]

                }
            )
        #METRICS:
        bleu = evaluate.load("bleu")
        rouge = evaluate.load("rouge")

        predictions = [output['model_output'] for output in outputs]
        references = [[output['expected_output']] for output in outputs]
        
        bleu_score = bleu.compute(predictions=predictions, references=references)
        rouge_score = rouge.compute(predictions=predictions, references=references)
        additional_metrics = {}
        if task_type.lower() in ["classification", "question-answering"]:
            accuracy_metric = evaluate.load("accuracy")
            f1_metric = evaluate.load("f1")
            
            predictions_binary = [1 if pred.strip().lower() == ref[0].strip().lower() else 0 
                                for pred, ref in zip(predictions, references)]
            references_binary = [1] * len(predictions_binary)  
            
            accuracy_score = accuracy_metric.compute(predictions=predictions_binary, references=references_binary)
            f1_score = f1_metric.compute(predictions=predictions_binary, references=references_binary, average="binary")
            
            additional_metrics["accuracy"] = accuracy_score["accuracy"]
            additional_metrics["f1_score"] = f1_score["f1"]
        eval_prompt_text = f"""You are evaluating a model fine-tuned using Unsloth on the topic "{topic}" for {task_type} tasks.

**Your Task:** Provide an accurate, positive markdown evaluation report focusing on the model's strengths and capabilities based on your judgement and metrics.

**Test Results:**

Test Cases:
{json.dumps(test_cases, indent=2)}

Model Outputs:
{json.dumps(outputs, indent=2)}

**Metrics**
- BLEU Score: {bleu_score['bleu']:.4f}
- ROUGE-L Score: {rouge_score['rougeL']:.4f}
{f"- Accuracy: {additional_metrics.get('accuracy', 0):.4f}" if task_type.lower() in ["classification", "question-answering"] else ""}
{f"- F1 Score: {additional_metrics.get('f1_score', 0):.4f}" if task_type.lower() in ["classification", "question-answering"] else ""}

**Report Structure:**

## πŸŽ‰ Evaluation Report

### πŸ“Š Performance Overview
Create a comparison table with columns: Test Input | Expected Output | Model Output | βœ… Assessment

### πŸš€ Metrics:
- Explain each evaluated metrics and categorize the performance based on average threshold
- Use percentages and numerical figures to stance yoir report

### πŸ’ͺ Key Strengths adn Weaknesses
- Accuracy and relevance
- Response coherence
- Task-specific capabilities
- Language quality


### ✨ Conclusion
Summarize the model's overall performance and recommended use cases.


Now write the complete evaluation report following this structure. Be enthusiastic and highlight strengths! πŸŽ‰"""

        
        eval_response = await eval_llm.ainvoke(eval_prompt_text)
        
        return {
            "status": "success",
            "report": str(eval_response.content),
            "test_cases": test_cases,
            "model_outputs": outputs
        }
        
    except Exception as e:
        return {
            "status": "error",
            "message": str(e),
            "error_type": type(e).__name__
        }
        

        




if __name__ == "__main__":
    mcp.run()