42 lines
15 KiB
JSON
42 lines
15 KiB
JSON
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 19, "axis": "style", "rep": 0, "err": null, "usage": {"prompt_tokens": 8427, "completion_tokens": 1897, "total_tokens": 10324, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1728, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.005901}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 19, "axis": "style", "rep": 0, "err": null, "usage": {"prompt_tokens": 11626, "completion_tokens": 5045, "total_tokens": 16671, "completion_tokens_details": {"reasoning_tokens": 4851}, "finish_reason": "stop"}, "cost": 0.031225}
|
|
{"model": "gemini-3.1-pro-preview", "judge": "gemini-3.1-pro-preview", "keyenv": "GEMINI_API_KEY", "ch": 19, "axis": "style", "rep": 0, "err": null, "usage": {"completion_tokens": 271, "prompt_tokens": 8303, "total_tokens": 10359, "finish_reason": "stop"}, "cost": 0.041278}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 19, "axis": "style", "rep": 1, "err": null, "usage": {"prompt_tokens": 8427, "completion_tokens": 1796, "total_tokens": 10223, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1600, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.005699}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 19, "axis": "style", "rep": 1, "err": null, "usage": {"prompt_tokens": 11626, "completion_tokens": 6035, "total_tokens": 17661, "completion_tokens_details": {"reasoning_tokens": 5806}, "finish_reason": "stop"}, "cost": 0.035185}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 19, "axis": "style", "rep": 2, "err": null, "usage": {"prompt_tokens": 8427, "completion_tokens": 2394, "total_tokens": 10821, "prompt_tokens_details": {"cached_tokens": 8320, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1984, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.006895}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 19, "axis": "style", "rep": 2, "err": null, "usage": {"prompt_tokens": 11626, "completion_tokens": 8490, "total_tokens": 20116, "cached_tokens": 10240, "completion_tokens_details": {"reasoning_tokens": 8239}, "prompt_tokens_details": {"cached_tokens": 10240}, "finish_reason": "stop"}, "cost": 0.045005}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 19, "axis": "fidelity", "rep": 0, "err": null, "usage": {"prompt_tokens": 8428, "completion_tokens": 2727, "total_tokens": 11155, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 2432, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.007561}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 19, "axis": "fidelity", "rep": 0, "err": null, "usage": {"prompt_tokens": 11633, "completion_tokens": 15151, "total_tokens": 26784, "completion_tokens_details": {"reasoning_tokens": 14955}, "finish_reason": "stop"}, "cost": 0.071655}
|
|
{"model": "gemini-3.1-pro-preview", "judge": "gemini-3.1-pro-preview", "keyenv": "GEMINI_API_KEY", "ch": 19, "axis": "fidelity", "rep": 0, "err": null, "usage": {"completion_tokens": 238, "prompt_tokens": 8304, "total_tokens": 14300, "finish_reason": "length"}, "cost": 0.08856}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 19, "axis": "fidelity", "rep": 1, "err": null, "usage": {"prompt_tokens": 8428, "completion_tokens": 2298, "total_tokens": 10726, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1920, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.006703}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 19, "axis": "fidelity", "rep": 1, "err": null, "usage": {"prompt_tokens": 11633, "completion_tokens": 6600, "total_tokens": 18233, "completion_tokens_details": {"reasoning_tokens": 6394}, "finish_reason": "stop"}, "cost": 0.037451}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 19, "axis": "fidelity", "rep": 2, "err": null, "usage": {"prompt_tokens": 8428, "completion_tokens": 2412, "total_tokens": 10840, "prompt_tokens_details": {"cached_tokens": 2816, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 2240, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.006931}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 19, "axis": "fidelity", "rep": 2, "err": null, "usage": {"prompt_tokens": 11633, "completion_tokens": 12602, "total_tokens": 24235, "cached_tokens": 2048, "completion_tokens_details": {"reasoning_tokens": 12434}, "prompt_tokens_details": {"cached_tokens": 2048}, "finish_reason": "stop"}, "cost": 0.061459}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 17, "axis": "style", "rep": 0, "err": null, "usage": {"prompt_tokens": 10467, "completion_tokens": 2687, "total_tokens": 13154, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 2304, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.007991}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 17, "axis": "style", "rep": 0, "err": null, "usage": {"prompt_tokens": 14843, "completion_tokens": 9268, "total_tokens": 24111, "completion_tokens_details": {"reasoning_tokens": 9019}, "finish_reason": "stop"}, "cost": 0.051173}
|
|
{"model": "gemini-3.1-pro-preview", "judge": "gemini-3.1-pro-preview", "keyenv": "GEMINI_API_KEY", "ch": 17, "axis": "style", "rep": 0, "err": null, "usage": {"completion_tokens": 192, "prompt_tokens": 10327, "total_tokens": 13466, "finish_reason": "stop"}, "cost": 0.058322}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 17, "axis": "style", "rep": 1, "err": null, "usage": {"prompt_tokens": 10467, "completion_tokens": 2199, "total_tokens": 12666, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1792, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.007015}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 17, "axis": "style", "rep": 1, "err": null, "usage": {"prompt_tokens": 14843, "completion_tokens": 12151, "total_tokens": 26994, "completion_tokens_details": {"reasoning_tokens": 11948}, "finish_reason": "stop"}, "cost": 0.062705}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 17, "axis": "style", "rep": 2, "err": null, "usage": {"prompt_tokens": 10467, "completion_tokens": 2566, "total_tokens": 13033, "prompt_tokens_details": {"cached_tokens": 10368, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 2112, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.007749}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 17, "axis": "style", "rep": 2, "err": null, "usage": {"prompt_tokens": 14843, "completion_tokens": 6129, "total_tokens": 20972, "cached_tokens": 14336, "completion_tokens_details": {"reasoning_tokens": 5957}, "prompt_tokens_details": {"cached_tokens": 14336}, "finish_reason": "stop"}, "cost": 0.038617}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 17, "axis": "fidelity", "rep": 0, "err": null, "usage": {"prompt_tokens": 10468, "completion_tokens": 2272, "total_tokens": 12740, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 2112, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.007161}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 17, "axis": "fidelity", "rep": 0, "err": "transport|TimeoutError|The read operation timed out", "usage": {}, "cost": 0.0}
|
|
{"model": "gemini-3.1-pro-preview", "judge": "gemini-3.1-pro-preview", "keyenv": "GEMINI_API_KEY", "ch": 17, "axis": "fidelity", "rep": 0, "err": null, "usage": {"completion_tokens": 237, "prompt_tokens": 10328, "total_tokens": 16324, "finish_reason": "length"}, "cost": 0.092608}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 17, "axis": "fidelity", "rep": 1, "err": null, "usage": {"prompt_tokens": 10468, "completion_tokens": 1811, "total_tokens": 12279, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1664, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.006239}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 17, "axis": "fidelity", "rep": 1, "err": "empty|finish=length|reasoning=True", "usage": {"prompt_tokens": 14850, "completion_tokens": 16000, "total_tokens": 30850, "completion_tokens_details": {"reasoning_tokens": 15999}, "finish_reason": "length"}, "cost": 0.078107}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 17, "axis": "fidelity", "rep": 2, "err": null, "usage": {"prompt_tokens": 10468, "completion_tokens": 3188, "total_tokens": 13656, "prompt_tokens_details": {"cached_tokens": 10368, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 3008, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.008993}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 17, "axis": "fidelity", "rep": 2, "err": null, "usage": {"prompt_tokens": 14850, "completion_tokens": 8887, "total_tokens": 23737, "cached_tokens": 14336, "completion_tokens_details": {"reasoning_tokens": 8750}, "prompt_tokens_details": {"cached_tokens": 14336}, "finish_reason": "stop"}, "cost": 0.049655}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 18, "axis": "style", "rep": 0, "err": null, "usage": {"prompt_tokens": 13502, "completion_tokens": 2471, "total_tokens": 15973, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 2304, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.008318}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 18, "axis": "style", "rep": 0, "err": null, "usage": {"prompt_tokens": 18666, "completion_tokens": 7613, "total_tokens": 26279, "completion_tokens_details": {"reasoning_tokens": 7367}, "finish_reason": "stop"}, "cost": 0.048185}
|
|
{"model": "gemini-3.1-pro-preview", "judge": "gemini-3.1-pro-preview", "keyenv": "GEMINI_API_KEY", "ch": 18, "axis": "style", "rep": 0, "err": null, "usage": {"completion_tokens": 314, "prompt_tokens": 13469, "total_tokens": 17486, "finish_reason": "stop"}, "cost": 0.075142}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 18, "axis": "style", "rep": 1, "err": null, "usage": {"prompt_tokens": 13502, "completion_tokens": 1770, "total_tokens": 15272, "prompt_tokens_details": {"cached_tokens": 13440, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1344, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.006915}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 18, "axis": "style", "rep": 1, "err": null, "usage": {"prompt_tokens": 18666, "completion_tokens": 5097, "total_tokens": 23763, "cached_tokens": 18432, "completion_tokens_details": {"reasoning_tokens": 4902}, "prompt_tokens_details": {"cached_tokens": 18432}, "finish_reason": "stop"}, "cost": 0.038121}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 18, "axis": "style", "rep": 2, "err": null, "usage": {"prompt_tokens": 13502, "completion_tokens": 1782, "total_tokens": 15284, "prompt_tokens_details": {"cached_tokens": 13440, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1600, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.00694}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 18, "axis": "style", "rep": 2, "err": null, "usage": {"prompt_tokens": 18666, "completion_tokens": 7766, "total_tokens": 26432, "cached_tokens": 18432, "completion_tokens_details": {"reasoning_tokens": 7597}, "prompt_tokens_details": {"cached_tokens": 18432}, "finish_reason": "stop"}, "cost": 0.048797}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 18, "axis": "fidelity", "rep": 0, "err": null, "usage": {"prompt_tokens": 13503, "completion_tokens": 2467, "total_tokens": 15970, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 2304, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.00831}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 18, "axis": "fidelity", "rep": 0, "err": null, "usage": {"prompt_tokens": 18673, "completion_tokens": 13217, "total_tokens": 31890, "completion_tokens_details": {"reasoning_tokens": 13036}, "finish_reason": "stop"}, "cost": 0.070607}
|
|
{"model": "gemini-3.1-pro-preview", "judge": "gemini-3.1-pro-preview", "keyenv": "GEMINI_API_KEY", "ch": 18, "axis": "fidelity", "rep": 0, "err": null, "usage": {"completion_tokens": 234, "prompt_tokens": 13470, "total_tokens": 19465, "finish_reason": "length"}, "cost": 0.09888}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 18, "axis": "fidelity", "rep": 1, "err": null, "usage": {"prompt_tokens": 13503, "completion_tokens": 3198, "total_tokens": 16701, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 2880, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.009772}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 18, "axis": "fidelity", "rep": 1, "err": null, "usage": {"prompt_tokens": 18673, "completion_tokens": 14412, "total_tokens": 33085, "completion_tokens_details": {"reasoning_tokens": 14221}, "finish_reason": "stop"}, "cost": 0.075387}
|
|
{"model": "gpt-5-mini", "judge": "gpt-5-mini", "keyenv": "OPENAI_API_KEY", "ch": 18, "axis": "fidelity", "rep": 2, "err": null, "usage": {"prompt_tokens": 13503, "completion_tokens": 1787, "total_tokens": 15290, "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0}, "completion_tokens_details": {"reasoning_tokens": 1600, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0}, "finish_reason": "stop"}, "cost": 0.00695}
|
|
{"model": "kimi-k2.6", "judge": "kimi-k2.6", "keyenv": "KIMI_API_KEY", "ch": 18, "axis": "fidelity", "rep": 2, "err": "transport|TimeoutError|The read operation timed out", "usage": {}, "cost": 0.0}
|