  e88 88e                               d8     
 d888 888b  8888 8888  ,"Y88b 888 8e   d88     
C8888 8888D 8888 8888 "8" 888 888 88b d88888   
 Y888 888P  Y888 888P ,ee 888 888 888  888     
  "88 88"    "88 88"  "88 888 888 888  888     
      b                                        
      8b,                                      
 
  e88'Y88                  d8           888    
 d888  'Y  ,"Y88b 888,8,  d88    ,e e,  888    
C8888     "8" 888 888 "  d88888 d88 88b 888    
 Y888  ,d ,ee 888 888     888   888   , 888    
  "88,d88 "88 888 888     888    "YeeP" 888    
                                               
PROUDLY PRESENTS

mini-magnum-12b-v1.1-exl2-longcal

Quantized using 115 rows of 8192 tokens from the default ExLlamav2-calibration dataset.

Branches:

main -- measurement.json
8b8h -- 8bpw, 8bit lm_head
6b8h -- 6bpw, 8bit lm_head
4b6h -- 4bpw, 6bit lm_head
2.25b6h -- 2.25bpw, 6bit lm_head

Original model link: intervitens/mini-magnum-12b-v1.1

Quanter's notes

As apparently the default dataset is supposed to be better in nearly all situations, I decided to start quanting using that in addition to my standard rpcal-fare. I'd appreciate real-world tests to confirm the hypothesis, though, so please leave a comment if you find this mode of quanting better than rpcal.

Original model README below.

This model is the miniature version of alpindale/magnum-72b-v1, a second entry 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 Mistral-Nemo-Base-2407. A new general purpose instruction dataset by kalomaze was added to the training mix for better coherence and general alignment. We are working on improving our dataset and training procedures, so expect new versions to come out soon.

Prompting

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

"""[INST] Hi there! [/INST]Nice to meet you!</s>[INST] Can I ask a question? [/INST]
"""

Credits

This model has been a team effort, credits go to:

Sao10K and kalomaze for help with (and cleaning up!) the dataset.
alpindale for the training.
Various other people for their continued help as we tuned the parameters, restarted failed runs. In no particular order: Doctor Shotgun, Lucy, Nopm, Mango, Intervitens, and the rest of the Silly Tilly.

Safety

...