alibayram/qwen3.5-toolcalling-v2

{
  "messages": [
    {
      "role": "system",
      "content": "You are a deep thinking AI...\n<tools>\n{\"name\": \"search_web\", \"description\": \"Search the web\", ...}\n</tools>"
    },
    {
      "role": "user",
      "content": "Find the latest news about AI developments"
    },
    {
      "role": "assistant",
      "content": "<think>\nUser wants latest AI news. I'll use search_web.\n</think>\n<tool_call>\n{\"name\": \"search_web\", \"arguments\": {\"query\": \"latest AI developments 2025\"}}\n</tool_call>"
    },
    {
      "role": "user",
      "content": "<tool_response>\n[{\"title\": \"New LLM achieves SOTA\", \"url\": \"...\"}]\n</tool_response>"
    },
    {
      "role": "assistant",
      "content": "Here are the latest AI developments: New LLM achieves state-of-the-art performance..."
    }
  ]
}

{
  "messages": [
    {
      "role": "system",
      "content": "You are an intelligent data science assistant with access to a stateful Jupyter notebook. You can use the add_and_execute_jupyter_code_cell tool.\n\nYou have access to: winequality-red.csv"
    },
    {
      "role": "user",
      "content": "What is the average alcohol content in the dataset?"
    },
    {
      "role": "assistant",
      "content": "<think>\nI need to load the CSV and compute the mean of the alcohol column.\n</think>\n{\"name\": \"add_and_execute_jupyter_code_cell\", \"arguments\": {\"code\": \"import pandas as pd\\ndf = pd.read_csv('/home/user/input/winequality-red.csv')\\nprint(df['alcohol'].mean())\"}}"
    },
    {
      "role": "assistant",
      "content": "10.422983114446529"
    },
    {
      "role": "assistant",
      "content": "<answer>\nThe average alcohol content is approximately 10.42%.\n</answer>"
    }
  ]
}

# --- UIGEN-X (ShareGPT) ---
role_map = {"human": "user", "gpt": "assistant", "system": "system"}

# --- Jupyter Agent (native messages) ---
# reasoning    -> assistant (merged with following tool_call)
# tool_call    -> assistant
# tool_output  -> user
# answer       -> assistant

from datasets import load_dataset

dataset = load_dataset("Mustafaege/qwen3.5-toolcalling-v2")

# Check both types of examples
for sample in dataset['train']:
    msgs = sample['messages']
    has_jupyter = any('jupyter_code_cell' in str(m['content']) for m in msgs)
    if has_jupyter:
        print("Jupyter agent example found!")
        break

from unsloth import FastLanguageModel
from trl import SFTTrainer, SFTConfig

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-1.7B",
    max_seq_length = 8192,  # Longer for multi-step reasoning
    load_in_4bit = True,
)

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset['train'],
    args = SFTConfig(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 8,
        max_seq_length = 8192,
    ),
)
trainer.train()