books/gu-zhenren/exp16/length_sweep.json

365 lines
No EOL
7.2 KiB
JSON

{
"by_finish": {
"stop": 413,
"length": 32,
"(none)": 32
},
"by_model_finish": {
"glm-5": {
"stop": 91
},
"gpt-5.4": {
"stop": 30
},
"grok-4.3": {
"stop": 25
},
"gemini-3.1-pro-preview": {
"length": 11,
"stop": 13,
"(none)": 2
},
"gpt-5-mini": {
"stop": 109,
"length": 1
},
"kimi-k2.6": {
"stop": 90,
"(none)": 29,
"length": 19
},
"deepseek-v4-flash": {
"stop": 19
},
"deepseek-v4-pro": {
"stop": 9,
"(none)": 1,
"length": 1
},
"mistral-large-2512": {
"stop": 27
}
},
"length_records": [
{
"ledger": "exp14/costs.jsonl",
"arm": "G-base",
"model": "gemini-3.1-pro-preview",
"id": "7.0",
"attempt": null,
"err": null,
"completion_tokens": 1061,
"cost": 0.10308
},
{
"ledger": "exp14/costs.jsonl",
"arm": "G-base",
"model": "gemini-3.1-pro-preview",
"id": "8.0",
"attempt": null,
"err": null,
"completion_tokens": 377,
"cost": 0.102774
},
{
"ledger": "exp14/costs.jsonl",
"arm": "G-base",
"model": "gemini-3.1-pro-preview",
"id": "10.0",
"attempt": null,
"err": null,
"completion_tokens": 319,
"cost": 0.103162
},
{
"ledger": "exp14/costs.jsonl",
"arm": "G-disc",
"model": "gemini-3.1-pro-preview",
"id": "7.0",
"attempt": null,
"err": null,
"completion_tokens": 317,
"cost": 0.10308
},
{
"ledger": "exp14/costs.jsonl",
"arm": "G-disc",
"model": "gemini-3.1-pro-preview",
"id": "8.0",
"attempt": null,
"err": null,
"completion_tokens": 1651,
"cost": 0.102774
},
{
"ledger": "exp14/costs.jsonl",
"arm": "G-disc",
"model": "gemini-3.1-pro-preview",
"id": "8.1",
"attempt": null,
"err": null,
"completion_tokens": 646,
"cost": 0.102314
},
{
"ledger": "exp14/costs.jsonl",
"arm": "G-disc",
"model": "gemini-3.1-pro-preview",
"id": "24.0",
"attempt": null,
"err": null,
"completion_tokens": 317,
"cost": 0.102604
},
{
"ledger": "exp14/costs.jsonl",
"arm": "G-disc",
"model": "gemini-3.1-pro-preview",
"id": "24.1",
"attempt": null,
"err": null,
"completion_tokens": 319,
"cost": 0.101642
},
{
"ledger": "exp14/rank_costs.jsonl",
"arm": null,
"model": "gemini-3.1-pro-preview",
"id": null,
"attempt": null,
"err": null,
"completion_tokens": 238,
"cost": 0.08856
},
{
"ledger": "exp14/rank_costs.jsonl",
"arm": null,
"model": "gemini-3.1-pro-preview",
"id": null,
"attempt": null,
"err": null,
"completion_tokens": 237,
"cost": 0.092608
},
{
"ledger": "exp14/rank_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.078107
},
{
"ledger": "exp14/rank_costs.jsonl",
"arm": null,
"model": "gemini-3.1-pro-preview",
"id": null,
"attempt": null,
"err": null,
"completion_tokens": 234,
"cost": 0.09888
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "gpt-5-mini",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=False",
"completion_tokens": 4000,
"cost": 0.008678
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.07015
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.069908
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.07026
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.06909
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.070172
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.070432
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": null,
"completion_tokens": 16000,
"cost": 0.068787
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.069911
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.070014
},
{
"ledger": "exp14/regate_costs.jsonl",
"arm": null,
"model": "kimi-k2.6",
"id": null,
"attempt": null,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.06911
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "D",
"model": "deepseek-v4-pro",
"id": "2.1",
"attempt": 1,
"err": null,
"completion_tokens": 8000,
"cost": 0.008481
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "K",
"model": "kimi-k2.6",
"id": "10.1",
"attempt": 1,
"err": null,
"completion_tokens": 16000,
"cost": 0.067627
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "K",
"model": "kimi-k2.6",
"id": "11.1",
"attempt": 1,
"err": null,
"completion_tokens": 16000,
"cost": 0.067858
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "K",
"model": "kimi-k2.6",
"id": "16.0",
"attempt": 1,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.068966
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "K",
"model": "kimi-k2.6",
"id": "17.0",
"attempt": 1,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.068826
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "K",
"model": "kimi-k2.6",
"id": "19.0",
"attempt": 1,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.068969
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "K",
"model": "kimi-k2.6",
"id": "2.1",
"attempt": 1,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.067913
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "K",
"model": "kimi-k2.6",
"id": "7.0",
"attempt": 1,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.068979
},
{
"ledger": "exp14b/costs.jsonl",
"arm": "K",
"model": "kimi-k2.6",
"id": "8.1",
"attempt": 1,
"err": "empty|finish=length|reasoning=True",
"completion_tokens": 16000,
"cost": 0.068465
}
]
}