MeTask-Rain's picture
Serving package (same as GitHub chenyuwang166-oss/MetaDecision-31B-Rain@6b8ed2d0)
068cb14 verified
Raw History Blame Contribute Delete
10 kB
{
"bench": {
"concurrency": [
1
],
"expected_public_count": 231,
"latency_offset_seconds": 0.15,
"latency_scale": 2.0,
"long_tail": {
"count": 23,
"projection_budgets": [
8192
],
"prompt_tokens": 79000,
"total_decisions": 1624
},
"original_types": false,
"output_prefix": "31b-exploration",
"target_slow_share": 0.49,
"timeout_seconds": 120
},
"calibration_file": "calibration.json",
"generation": {
"bad_words_ids": null,
"begin_suppress_tokens": null,
"do_sample": false,
"forced_bos_token_id": null,
"forced_decoder_ids": null,
"forced_eos_token_id": null,
"length_penalty": 1.0,
"max_new_tokens": 192,
"min_length": 0,
"min_new_tokens": 0,
"no_repeat_ngram_size": 0,
"reasoning_words": 50,
"renormalize_logits": false,
"repetition_penalty": 1.0,
"seed": 0,
"suppress_tokens": null,
"temperature": 1.0,
"top_k": 0,
"top_p": 1.0,
"wall_timeout_seconds": 60
},
"generation_reference": {
"intended_overrides": {
"do_sample": {
"reason": "Frozen greedy decoding; checkpoint sampling policy is intentionally overridden.",
"value": false
},
"suppress_tokens": {
"reason": "All reported measurements were taken without checkpoint token suppression; drafts only, fast-path candidate scoring unaffected.",
"value": null
},
"temperature": {
"reason": "Frozen greedy decoding; checkpoint sampling policy is intentionally overridden.",
"value": 1.0
},
"top_k": {
"reason": "Frozen greedy decoding; checkpoint sampling policy is intentionally overridden.",
"value": 0
},
"top_p": {
"reason": "Frozen greedy decoding; checkpoint sampling policy is intentionally overridden.",
"value": 1.0
}
}
},
"interface": {
"boolean_keys": "request"
},
"item_deadline": {
"chunk_tokens": 32,
"ewma_alpha": 0.3,
"item_seconds": 9.0,
"min_chunk_tokens": 8,
"readout_reserve_init_s": 1.0,
"readout_reserve_min_s": 0.3,
"readout_safety": 1.5,
"response_reserve_s": 0.35,
"token_safety": 1.25,
"warmup_repeats": 3,
"warmup_tokens": 33
},
"item_rules": {
"reread_tau": 4.0,
"slow_types": [
"choice",
"noul",
"score"
]
},
"limits": {
"compression_reserve_tokens": 32,
"cost_per_1000_decisions_usd": 0.0646,
"estimated_chars_per_token": 4,
"max_prompt_tokens": 6653,
"median_latency_seconds": 1.0,
"oom_retry_prompt_tokens": 4096,
"physical_max_prompt_tokens": 6653,
"precompression_trigger_chars_per_token": 8,
"raw_median_latency_seconds": 0.425,
"state_chunk_tokens": 200,
"state_head_tokens": 512,
"state_tail_tokens": 512,
"truncation_head_frac": 0.6
},
"model": {
"adapter": "model-subdir:adapter",
"attn_implementation": "sdpa",
"backend": "vllm",
"device_map": {
"": 0
},
"enable_prefix_caching": true,
"model_key": "metask-gemma4-31b-lora-w4a16",
"name": "MetaDecision-31B-Rain",
"path": "",
"quant": "awq",
"version": "MetaDecision-31B-Rain",
"vllm_expected_version": "0.31.0",
"vllm_gpu_memory_utilization": 0.94,
"vllm_max_model_len": 6656,
"vllm_options": {
"gpu_memory_utilization": 0.94,
"max_lora_rank": 16,
"max_model_len": 6656,
"max_num_seqs": 1
}
},
"package": {
"finalized": true,
"nonformal_fitted": true,
"sample_only": false,
"version": "MetaDecision-31B-Rain"
},
"pipeline": {
"arm": "trained",
"model_identity_sha256": "0513d121db967ec1a6aaccf627d3c42149d092dd027298da5b53684b030821ea",
"protocol": "adapter_arm_31b_v3"
},
"post_calibration": {
"choice": {
"config": {
"debias": false,
"temperature": false
},
"params": {}
},
"noul": {
"config": {
"floor": null,
"temperature": "type"
},
"params": {
"T": 0.5
}
},
"score": {
"config": {
"mode_weight": 1.0,
"temperature": false
},
"params": {}
}
},
"pricing": {
"margin": 0.0,
"model_key": "31B",
"models": {
"31B": {
"input_per_million": 0.059819178473256074,
"output_per_million": 0.059819178473256074
}
},
"submission_models": {
"MetaDecision-31B-Rain": "31B",
"metask-gemma4-31b-lora-w4a16": "31B",
"metask-jev-rain-31B-trained": "31B"
}
},
"profiles": {
"metask-gemma4-31b-lora-w4a16": {
"bench": {
"target_slow_share": 0.49
},
"generation": {
"bad_words_ids": null,
"begin_suppress_tokens": null,
"do_sample": false,
"forced_bos_token_id": null,
"forced_decoder_ids": null,
"forced_eos_token_id": null,
"length_penalty": 1.0,
"max_new_tokens": 192,
"min_length": 0,
"min_new_tokens": 0,
"no_repeat_ngram_size": 0,
"reasoning_words": 50,
"renormalize_logits": false,
"repetition_penalty": 1.0,
"seed": 0,
"suppress_tokens": null,
"temperature": 1.0,
"top_k": 0,
"top_p": 1.0,
"wall_timeout_seconds": 60
},
"limits": {
"max_prompt_tokens": 6653,
"physical_max_prompt_tokens": 6653
},
"model": {
"backend": "vllm",
"name": "MetaDecision-31B-Rain",
"quant": "awq"
},
"pricing": {
"model_key": "31B"
},
"routing": {
"cost_fuse": {
"arm": 0.9,
"enabled": true,
"release": 0.85,
"warmup": 100
},
"fuse": 0.49,
"long_item_slow_policy": "fast_only",
"margin_kind": "logratio",
"margin_threshold": 6.625023671062799,
"margin_threshold_fitted_for": {
"achieved_share": 0.5207547169811321,
"adapter": "model-subdir:adapter",
"backend": "vllm",
"fitted_on": "our own independently built development set (fit split): fast-path margins of the earlier service routing, eligible_items rule (the lowest-margin share of the decisions that can take the drafting pass)",
"margin_kind": "logratio",
"max_new_tokens": 192,
"max_prompt_tokens": 6653,
"model_key": "metask-gemma4-31b-lora-w4a16",
"n": 530,
"placeholder": false,
"prompt_sha": "e18ed7a06eb52f1b351458f384710a2c240aaf2b4f64d916b776c6d18ca3f938",
"quant": "awq",
"reasoning_words": 50,
"request_contract": "typesafe.criteria.v1"
},
"mode": "routed",
"quota": 0.49,
"slow_max_prompt_tokens": 840,
"target_slow_share": 0.49,
"tier": "normal",
"tiers": {
"hard": {
"fast_only": false,
"max_new_tokens": 192,
"reasoning_words": 50
},
"long": {
"fast_only": true,
"max_prompt_tokens": 6653
},
"normal": {
"fast_only": false,
"max_prompt_tokens": 6653
}
},
"unparsed_policy": "force_answer"
}
}
},
"prompt": {
"answer_marker": "ANSWER:",
"cot_instructions": {
"choice": "First reason step by step in at most {reasoning_words} words, tracking any intermediate state explicitly. Then output a final line `ANSWER: <letter>`.",
"noul": "First reason step by step in at most {reasoning_words} words, tracking any intermediate state explicitly. Then output a final line of the form `ANSWER: yes` or `ANSWER: no`.",
"score": "First reason step by step in at most {reasoning_words} words, tracking any intermediate state explicitly. Then output a final line `ANSWER: <number>`."
},
"enable_thinking": false,
"system": "You are a careful, decisive judge. Read the context, apply every rule, definition, threshold, exception and date it contains, and answer the question. Reply with exactly one word and nothing else."
},
"routing": {
"cost_fuse": {
"arm": 0.9,
"enabled": true,
"release": 0.85,
"warmup": 100
},
"fuse": 0.49,
"long_item_slow_policy": "fast_only",
"margin_kind": "logratio",
"margin_threshold": 6.625023671062799,
"margin_threshold_fitted_for": {
"achieved_share": 0.5207547169811321,
"adapter": "model-subdir:adapter",
"backend": "vllm",
"fitted_on": "our own independently built development set (fit split): fast-path margins of the earlier service routing, eligible_items rule (the lowest-margin share of the decisions that can take the drafting pass)",
"margin_kind": "logratio",
"max_new_tokens": 192,
"max_prompt_tokens": 6653,
"model_key": "metask-gemma4-31b-lora-w4a16",
"n": 530,
"placeholder": false,
"prompt_sha": "e18ed7a06eb52f1b351458f384710a2c240aaf2b4f64d916b776c6d18ca3f938",
"quant": "awq",
"reasoning_words": 50,
"request_contract": "typesafe.criteria.v1"
},
"mode": "routed",
"quota": 0.49,
"resume": false,
"run_id": null,
"slow_max_prompt_tokens": 840,
"state_dir": null,
"target_slow_share": 0.49,
"tier": "normal",
"tiers": {
"hard": {
"fast_only": false,
"max_new_tokens": 192,
"reasoning_words": 50
},
"long": {
"fast_only": true,
"max_prompt_tokens": 6653
},
"normal": {
"fast_only": false,
"max_prompt_tokens": 6653
}
},
"unparsed_policy": "force_answer"
},
"server": {
"bind": "127.0.0.1",
"max_questions": 64,
"max_request_bytes": 16777216,
"port": 8000,
"socket_timeout_seconds": 120,
"warmup": true,
"warmup_slow_tokens": 2
},
"usage": {
"basis": "engine"
}
}