kaistai / selfee Goto Github PK

Hi,
Realy nice project. Any plan on releasing an uncensored version? Usually they have better performance.
Also a wiki with use-cases and how-to would be nice.
Also, an increase of the context window seems possible, see the latest projects released by theBloke on huggingface/ kaiokendev work.

Kudos,

Collecting torch==2.0.1 (from -r requirements.txt (line 4))
Downloading torch-2.0.1-cp311-cp311-manylinux1_x86_64.whl (619.9 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 619.9/619.9 MB 3.0 MB/s eta 0:00:00
Collecting flash-attn==0.2.8 (from -r requirements.txt (line 5))
Using cached flash_attn-0.2.8.tar.gz (1.8 MB)
Preparing metadata (setup.py) ... error
error: subprocess-exited-with-error

× python setup.py egg_info did not run successfully.
│ exit code: 1
╰─> [6 lines of output]
Traceback (most recent call last):
File "", line 2, in
File "", line 34, in
File "/tmp/pip-install-tjhjvhnh/flash-attn_d4c66fc1204a49cf9a44de9ebbba6e8d/setup.py", line 10, in
import torch
ModuleNotFoundError: No module named 'torch'
[end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
error: metadata-generation-failed

× Encountered error while generating package metadata.
╰─> See above for output.

note: This is an issue with the package mentioned above, not pip.
hint: See above for details

Got this error message:

NETWORK ERROR DUE TO HIGH TRAFFIC. PLEASE REGENERATE OR REFRESH THIS PAGE. (error_code: 4)

Letting the model provide self-feedback is a very interesting idea, I'm curious about how it would perform if training without feedback and only using the final revision answer. Have you tried that?

I didn't know how else to communicate my thoughts about the project, so I opened this issue.

The project is absolutely interesting and shows its fruits! but there is something particular I noticed, while testing it using your demo, I noticed that even if the model gave bad answers or was hallucinating, the feedback it gave itself was always very positive and the ratings are never lower than 8.

I checked the training dataset on the huggingface page, and noticed that there is also a large presence of positive feedback in the dataset!

So I thought, what if a lot of bad examples were also introduced into the dataset with the corresponding feedback misjudging the answer? Then the model should become more aware of what is right or wrong in the answers.

What do you think?

I really appreciate your work, a huge thank you!