license: cc-by-nc-4.0
tags:
- moe
- merge
pipeline_tag: text-generation
model-index:
- name: Proctora
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: AI2 Reasoning Challenge (25-Shot)
type: ai2_arc
config: ARC-Challenge
split: test
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 67.83
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Karko/Proctora
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (10-Shot)
type: hellaswag
split: validation
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 86.68
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Karko/Proctora
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU (5-Shot)
type: cais/mmlu
config: all
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 65.49
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Karko/Proctora
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (0-shot)
type: truthful_qa
config: multiple_choice
split: validation
args:
num_few_shot: 0
metrics:
- type: mc2
value: 59.55
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Karko/Proctora
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Winogrande (5-shot)
type: winogrande
config: winogrande_xl
split: validation
args:
num_few_shot: 5
metrics:
- type: acc
value: 79.79
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Karko/Proctora
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GSM8k (5-shot)
type: gsm8k
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 71.95
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Karko/Proctora
name: Open LLM Leaderboard
Proctora is a MoE model made of
- OpenPipe/mistral-ft-optimized-1227 as a base model
- SanjiWatsuki/Kunoichi-7B as a first expert dedicated to RP tasks.
- samir-fama/SamirGPT-v1 as a second expert for factual answers.
Being based on Mixtral architecture it has a natural context length of 32K, which is great.
On Openllm leaderboard it achieves a score of 71.88 which is interesting to some extent but does not really reflect the intented capacities of the model.
This model has been originally produced as a result of experimentations with mergekit. Then among my collection of LLMs, Proctora has been selected to be the "grader" in an AI-RPG evaluation suite that I am currently building. Indeed, it produced the intended grades according to given rubrics more often than other "higher performing" models in the leaderboard.
However, I also tested it in various RP scenarii using text-generation-webui (putting the character card in the system parameters and/or other world information), and I was quite impressed by the quality of the logic (relatively to other popular RP models). For example, it took in account special powers limitations better than other models. Or it managed curse activations and weaknesses better than other models that are about twice the size. Also when acting as the player (and the user being the game master), Proctora was not only able to play in character but also sometimes to make clever decision to achieve its objectives.
Having the excellent SanjiWatsuki/Kunoichi-7B as an expert, the model is uncensored. Use with caution.
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 71.88 |
| AI2 Reasoning Challenge (25-Shot) | 67.83 |
| HellaSwag (10-Shot) | 86.68 |
| MMLU (5-Shot) | 65.49 |
| TruthfulQA (0-shot) | 59.55 |
| Winogrande (5-shot) | 79.79 |
| GSM8k (5-shot) | 71.95 |