44 lines
1.9 KiB
JSON
44 lines
1.9 KiB
JSON
{
|
|
"traces": [],
|
|
"train": [],
|
|
"demos": [],
|
|
"signature": {
|
|
"instructions": "Judge whether this reply adequately addresses the original email.\n\nEvaluate across three dimensions:\n- Relevance: Does the reply directly address the main request or question\n in the original email?\n- Tone consistency: Is the reply's register (formal/casual) consistent\n with the original email?\n- Conciseness: Does the reply say what needs to be said without being\n evasive or padded?\n\nOutput 'good' if the reply is adequate or better on all three dimensions.\nOutput 'poor' if the reply fails to address the original, is tonally\ninappropriate, or is evasive/uninformative.\n\nNote: judge based solely on the text provided. Missing attachments\n(referenced as << filename >>) should not by themselves cause a 'poor'\nrating.\n\nAs you reason, note any uncertainties, ambiguities, and gaps \u2014 both in\nhow the criteria apply and in whether you have enough context to judge\nconfidently.",
|
|
"fields": [
|
|
{
|
|
"prefix": "Original Email:",
|
|
"description": "The original email being replied to (may be truncated)."
|
|
},
|
|
{
|
|
"prefix": "Reply:",
|
|
"description": "The reply email body (may be truncated)."
|
|
},
|
|
{
|
|
"prefix": "Reasoning:",
|
|
"description": "Step-by-step reasoning, including uncertainty, ambiguity, or uncertainty in your thought process when relevant."
|
|
},
|
|
{
|
|
"prefix": "Quality:",
|
|
"description": "'good' if the reply adequately addresses the original, 'poor' otherwise."
|
|
}
|
|
]
|
|
},
|
|
"lm": {
|
|
"model": "together_ai/Qwen/Qwen2.5-7B-Instruct-Turbo",
|
|
"model_type": "chat",
|
|
"cache": true,
|
|
"num_retries": 3,
|
|
"finetuning_model": null,
|
|
"launch_kwargs": {},
|
|
"train_kwargs": {},
|
|
"temperature": null,
|
|
"max_tokens": null
|
|
},
|
|
"metadata": {
|
|
"dependency_versions": {
|
|
"python": "3.11",
|
|
"dspy": "3.1.3",
|
|
"cloudpickle": "3.1"
|
|
}
|
|
}
|
|
} |