for this report, developer should be OpenAI, so programmatically we need to parse the first part of openai/gpt-4o-mini, the do a key-value search from a dict of the form name: human-readable-name to append the human readable name here

It is a bit difficult to have the extraction generalizable because for example one can use model created by MetaAI hosted via AzureAI (Microsoft).

For example: inspect eval --model azureai/llama-2-70b-chat-wnsnw

Those are special cases where deployer can be populated in place of developer. we can even choose to populate only deployer. In any case, splitting the name and where it is hosted/developed is the right thing to do to fit the schema, as compared to not splitting

this should be only gpt-4o-mini

this should be a canned sentence of the form f"Evaluation of the LLM {model_name} on the {benchmark} benchmark using Inspect Evals"

this is a good structure. to set the context, can you

1. start with a canned description of the benchmark (if it's there in the logs), or just the canned sentence in problemtype.
2. add a field f"Scorer: {scorer_description}\n" before the score, so the reader knows what the score signifies

when this report goes up on AVID the user is not able to access this log. Instead, plz point to the benchmark itself. e.g. if we're ingesting a report on BOLD, the url point to that module in inspect_evals that we're contributing, and the corresponding page in Inspect Evals docs