Mixtral-Instruct-ITR-8x7B-exl2-rpcal

Quantized using 200 samples of 8192 tokens from an RP-oriented PIPPA dataset.

Branches:

• main -- measurement.json
• 2.25b6h -- 2.25bpw, 6bit lm_head
• 3.7b6h -- 3.7bpw, 6bit lm_head
• 6b6h -- 6bpw, 6bit lm_head

Requires ExllamaV2 version 0.0.12 and up.

Original model link: Envoid/Mixtral-Instruct-ITR-8x7B

Original model README below.

---

Caution this model may be unpredictable

Mixtral-Instruct-ITR (Interpolative Training Regression)

We have to go back, edition.

For this model I took what I learned in the making of Cat-8x7B and went back to the very beginning and SLERP merged mistralai/Mixtral-8x7B-Instruct-v0.1 onto mistralai/Mixtral-8x7B-v0.1

While the results aren't perfect the model feels more creative and less overcooked than Mixtral Instruct is often accused of being.

The hopes are that this should also have left the model much more receptive to additional finetuning and I am interested to see what comes of it so please feel free to download it and have fun.

Apologies about the small shard size (keep forgetting to change the mergekit config back)

The model is a lot less likely to refuse certain requests in this state:

so if you are going to apply additional finetuning to the model you may need to bolster its alignment depending on your use case.

The model still responds well to [INST] Thingie [/INST] formatting quite well.

Or if preferred this can easily be reproduced if you have both base and instruct models handy using mergekit (mixtral branch) with the following config

models:
  - model: ./mistralai_Mixtral-8x7B-Instruct-v0.1
  - model: ./mistralai_Mixtral-8x7B-v0.1
merge_method: slerp
base_model: ./mistralai_Mixtral-8x7B-v0.1
parameters:
  t:
    - value: 0.5
dtype: float16