Pengumpulan Data: Trajektori dan Pemutaran Ulang Jejak

Data Adalah Produk

Seperti Apa Trajektori Agen

trajectory = {
    'task': 'Refactor the auth module',
    'messages': [
        {'role': 'system', 'content': '...'},
        {'role': 'user', 'content': 'Task...'},
        {'role': 'assistant', 'content': 'Plan...', 'tool_calls': [...]},
        {'role': 'tool', 'content': '...'},
        ...
    ],
    'outcome': 'success'
}

Menambang Rekam Jejak Produksi

for trace in production_traces:
    if trace.outcome == 'success' and trace.user_feedback == 'thumbs_up':
        save_to_training_set(trace)

Pemutaran Ulang Rekam Jejak

def replay(trace):
    messages = trace.messages[:1]   # just the initial user msg
    for expected_msg in trace.messages[1:]:
        actual = agent.step(messages)
        if actual.role == 'assistant':
            messages.append(actual)
        elif expected_msg.role == 'tool':
            messages.append(expected_msg)   # replay tool result

Menyaring Rekam Jejak Berkualitas Tinggi

Distilasi dari Model Besar

for task in task_list:
    traj = big_model_agent.run(task)
    if traj.success:
        save_for_training(traj)

# Now fine-tune Llama 8B on the GPT-4o traces.

Memangkas Trajektori Buruk

Format untuk Penyetelan Halus

{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
{"messages": [...]}

Pemanggilan Alat dalam Data Pelatihan

{
  "messages": [
    {"role": "user", "content": "Weather in Paris?"},
    {"role": "assistant", "tool_calls": [{...}]},
    {"role": "tool", "tool_call_id": "...", "content": "{...}"},
    {"role": "assistant", "content": "It is 18C and sunny."}
  ]
}

Contoh Negatif

preferences = [
    {'prompt': '...', 'chosen': good_response, 'rejected': bad_response}
]

Kalkulator Ukuran Kumpulan Data

Membuat Versi Kumpulan Data

Kurasi dengan Keterlibatan Manusia

Sumber Rekam Jejak Terbaik

Ringkasan