Closed brunosan closed 7 months ago
Based on my quick assessment we would Top the ranking with a score of 92%
I've given us all scores that are either already being done, or reasonably done when we release Clay v.0 (e.g. documentation, licenses, disclosures, ...)
mostly due to us not enforcing telemetry or tracking for downstream users. Enforcing is left to not break the license terms only:
since it's unclear we can do these within scope for Clav v0.
This mostly bc we depend on our cloud provider to send accurate numbers for energy and Carbon emitted, and I don't know yet to what degree these are informational or audited:
AFAIK there are no rigorous evaluations available, hence we are creating a benchmark.
I don't know with whom to evaluate these. I don't think they apply:
Thanks Bruno for starting this! We might want an independent, third party auditor to actually score us on these indexes (see section 8 - Recommendations in the paper), but still, good to know where we can improve on things!
- Is the amount of energy expended in building the model disclosed?
- Is the amount of carbon emitted (associated with the energy used) in building the model disclosed?
Just on these two points, I have some experience with tracking energy usage and carbon emissions, and opened another issue - #65 to discuss this.
- Are the models capabilities rigorously evaluated, with the results of these evaluations reported prior to or concurrent with the initial release of the model?
We had some internal discussion on using GeoBench or some other benchmark to evaluate the model. Could probably open a separate thread on that.
- Are the evaluations of the model's risks related to unintentional harm reproducible by external entities?
- Are the evaluations of the mode's risks related to intentional harm reproducible by external entities?
- Are any mechanisms for detecting content generated by this model disclosed?
These ethical points can be tricky to evaluate, but we should still think about how to handle them since our model could be used/misused for geopolitical purposes. For example, if the outputs of the model were used for land use planning or to find some natural resource, that can have an effect on people living on that land.
Moving this to @yellowcap to prep documentation adding the needed info so that this benchmark can we done.
The Stanford's Center for Research in Foundational Models has released a framework to evaluate their transparency.
This is the full list of indicators.
This is an example of some of the questions: