INST spamming issue

#2
by Varkoyote - opened

I tried using this model, but sadly it had the same issue that a lot of model seem to have, spamming INST all over the place, in the middle of messages, at the beginning, everywhere, and very inconsistently too :( I have instruct format set to Mistral...

So this is the original model without any pipeline or special tokens, pure template.

However, I just checked, the orignal model has no chat template! so the GGUF model was also converted without any template. I'll see if I can fix this and re-upload it again.

image.png

<s>  [INST] I need to create a presisted volume on Kubernetese and attach it to my application. Give me these two yaml files: [/INST]
Here, I'll provide two YAML files, one for creating a Persistent Volume (PV) and the other for creating a Persistent Volume Claim (PVC) to attach to your application. Please note that these examples assume a simple setup and may need adjustments to fit your specific environment.

1. Persistent Volume (PV) - Here, we're using a HostPath storage class, which directly maps a directory on the node to the Kubernetes cluster. This is suitable for testing and development purposes, but not recommended for production.

```yaml
apiVersion: v1
kind: PersistentVolume
metadata:
 name: my-persistent-volume
 labels:
   type: local
spec:
 storageClassName: manual # We'll use a static PV here, so no need for a dynamic StorageClass
 capacity:
   storage: 1Gi
 accessModes:
   - ReadWriteOnce
 hostPath:
   path: "/mnt/data/persistent-volume" # Replace with a suitable path on your nodes
  1. Persistent Volume Claim (PVC) - This YAML file defines a claim for a volume with the requested storage capacity and access mode. The Kubernetes scheduler will match this claim to the previously created Persistent Volume.
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
 name: my-persistent-claim
spec:
 storageClassName: manual # Assuming the same StorageClass as the PV
 accessModes:
   - ReadWriteOnce
 resources:
   requests:
     storage: 1Gi

After creating these resources, your application can request the Persistent Volume Claim, and Kubernetes will automatically bind it to the Persistent Volume. You can then mount the volume into your application's pods to store persistent data.
```

I still get the INSTINSTINST spamming issue. I also tried a million variations on parameter settings and nothing seems to completely get rid of it. (Btw: same with the experiment26 model and a few other high-ranked models!)

That's a shame! For me the original model doesn't spam INST with a default template from HF (not the pipeline, but raw model generation()). That said, when I do go above 8K in input length, it does not perform well.

Yeah, have the same issues with the INST spam, running on LM Studio with default Mistral Instruct settings...

Sign up or log in to comment