luigi86/magnum-v2.5-12b-kto_mlx

MLX Format and Quantizations for Magnum v2.5 12b KTO

Converted uncompressed and tested using the mlx_lm utility on a 64GiB URAM M1 Max.

• 8bpw quants

• 16bpw model

See original model for further details.

Original Model card

v2.5 KTO is an experimental release; we are testing a hybrid reinforcement learning strategy of KTO + DPOP, using rejected data sampled from the original model as "rejected". For "chosen", we use data from the original finetuning dataset as "chosen". This was done on a limited portion of of primarily instruction following data; we plan to scale up a larger KTO dataset in the future for better generalization.

This is the 5th in a series of models designed to replicate the prose quality of the Claude 3 models, specifically Sonnet and Opus. This model is fine-tuned on top of anthracite-org/magnum-12b-v2.

Prompting

Model has been Instruct tuned with the ChatML formatting. A typical input would look like this:

"""<|im_start|>user
Hi there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
<|im_start|>user
Can I ask a question?<|im_end|>
<|im_start|>assistant
"""

Credits

• Stheno dataset (filtered)
• kalomaze/Opus_Instruct_25k
• Nopm/Opus_WritingStruct
• Gryphe/Sonnet3.5-SlimOrcaDedupCleaned (A ~16k rows subset)
• kalomaze/Opus_Instruct_3k

This model has been a team effort, and the credits goes to all members of Anthracite.

Safety

...