Add Cohere Command R7B, replace older Command R+ handler

[harry-cohere]

Greetings! It's been a while since our last contribution to BFCL, and the new versions and recent improvements are great to see.

This PR adds our latest model, released on Friday. I've also replaced our older models because it simplifies the code within the BFCL framework.

When I run this PR against 3245d9 I get the following results (without REST category sanity checks):

|Name |Overall Acc|Model      |Model Link                         |Cost ($ Per 1k Function Calls)|Latency Mean (s)|Latency Standard Deviation (s)|Latency 95th Percentile (s)|Non-Live AST Acc|Non-Live Simple AST|Non-Live Multiple AST|Non-Live Parallel AST|Non-Live Parallel Multiple AST|Non-Live Exec Acc|Non-Live Simple Exec|Non-Live Multiple Exec|Non-Live Parallel Exec|Non-Live Parallel Multiple Exec|Live Acc|Live Simple AST|Live Multiple AST|Live Parallel AST|Live Parallel Multiple AST|Multi Turn Acc|Multi Turn Base|Multi Turn Miss Func|Multi Turn Miss Param|Multi Turn Long Context|Relevance Detection|Irrelevance Detection|Organization|License     |
|-----|-----------|-----------|-----------------------------------|------------------------------|----------------|------------------------------|---------------------------|----------------|-------------------|---------------------|---------------------|------------------------------|-----------------|--------------------|----------------------|----------------------|-------------------------------|--------|---------------|-----------------|-----------------|--------------------------|--------------|---------------|--------------------|---------------------|-----------------------|-------------------|---------------------|------------|------------|
|Run 1|53.37%     |Command R7B|https://cohere.com/blog/command-r7b|0.1                           |5.22            |9.75                          |12.64                      |81.00%          |68.00%             |91.50%               |84.50%               |80.00%                        |81.21%           |84.86%              |90.00%                |80.00%                |70.00%                         |74.23%  |63.95%         |69.71%           |50.00%           |70.83%                    |5.12%         |7.00%          |1.00%               |6.50%                |6.00%                  |61.11%             |80.68%               |Cohere      |cc-by-nc-4.0|
|Run 2|53.59%     |Command R7B|https://cohere.com/blog/command-r7b|0.1                           |3.19            |10.43                         |6.29                       |81.21%          |68.33%             |91.50%               |84.50%               |80.50%                        |82.46%           |85.36%              |92.00%                |80.00%                |72.50%                         |74.28%  |63.18%         |69.71%           |50.00%           |83.33%                    |5.12%         |7.50%          |1.50%               |6.00%                |5.50%                  |61.11%             |80.47%               |Cohere      |cc-by-nc-4.0|

I notice some variance between runs but haven't concluded where this is coming from - I'm confident, however, that the model regularly scores the higher score.

Thanks in advance for reviewing this PR and I look forward to seeing us on the leaderboard!

[Fanjia-Yan]

Hi Cohere,

Congrats on the release and thanks for contributing the really nice tool calling interface to interact with command models! We have one question before onboarding r7b on the leaderboard:

We typically don't deprecate a model unless the next generation alternative releases. With that, would v2 client be able to support full tool usage capability of Command R+ and Command R now or in the future? We do like to see models of different sizes on leaderboard for better comparison :) Thank you.

cc: @HuanzhiMao @ShishirPatil

[harry-cohere]

Thanks @Fanjia-Yan ! All of our models support the newer version, so I've added command-r-plus back and it now uses the updated handler. When I run it against 3245d9 I get this score:

|Overall Acc|Model              |Model Link                                           |Cost ($ Per 1k Function Calls)|Latency Mean (s)|Latency Standard Deviation (s)|Latency 95th Percentile (s)|Non-Live AST Acc|Non-Live Simple AST|Non-Live Multiple AST|Non-Live Parallel AST|Non-Live Parallel Multiple AST|Non-Live Exec Acc|Non-Live Simple Exec|Non-Live Multiple Exec|Non-Live Parallel Exec|Non-Live Parallel Multiple Exec|Live Acc|Live Simple AST|Live Multiple AST|Live Parallel AST|Live Parallel Multiple AST|Multi Turn Acc|Multi Turn Base|Multi Turn Miss Func|Multi Turn Miss Param|Multi Turn Long Context|Relevance Detection|Irrelevance Detection|Organization |License     |
|-----------|-------------------|-----------------------------------------------------|------------------------------|----------------|------------------------------|---------------------------|----------------|-------------------|---------------------|---------------------|------------------------------|-----------------|--------------------|----------------------|----------------------|-------------------------------|--------|---------------|-----------------|-----------------|--------------------------|--------------|---------------|--------------------|---------------------|-----------------------|-------------------|---------------------|-------------|------------|
|51.17%     |Command-R-Plus (FC)|https://txt.cohere.com/command-r-plus-microsoft-azure|7.71                          |2.75            |9.26                          |4.49                       |76.56%          |72.25%             |89.50%               |82.00%               |62.50%                        |79.91%           |90.14%              |90.00%                |82.00%                |57.50%                         |63.26%  |73.64%         |71.79%           |43.75%           |58.33%                    |15.00%        |23.00%         |10.00%              |11.50%               |15.50%                 |94.44%             |50.57%               |Cohere For AI|cc-by-nc-4.0|

This is +5.1% from the leaderboard score, and the increase mainly comes from the multihop category.

Looking into the increase I think part of the reason is that there are bugs in the replaced handler which lowered the overall score. For example the base handler adds a holdout function message with role/content, but the replaced handler expects them as role/message, which would explain the 10% increase in the missed function category. There may be other sources of the increase, too.

Another change is the cost per 1K functions - I see an increase of around $2.30 per 1K functions. Is it possible that the price is increased because the multi turn categories now succeed in more cases?

[Fanjia-Yan]

Thank you Harry! That addresses all my concerns. And yes, higher cost per 1K functions typically means the model attempts more and have less early stop. @HuanzhiMao will review code change and merge that in.

[ShishirPatil]

LGTM. Thank you for the PR and congratulations on the new model launch @harry-cohere