point-alpaca

What is this?

This is released weights recreated from Stanford Alpaca, an experiment in fine-tuning LLaMA on a synthetic instruction dataset.

This is not LoRA, this is a full fine-tune for 3 epochs on 8x A100 80 GB, loss ≈2 ➔ ≈0.5.

Can I try this somewhere?

Yes! Announcement thread to our frontend where you can try the 7B: https://twitter.com/PointNetwork/status/1637178814210908160

Try it here: https://alpaca.point.space

What are hardware requirements to run it locally?

It takes 16 GB of VRAM unquantized, 8 GB of VRAM when 8-bit quantized (11 GB of normal RAM to load it).

It's confirmed that it can run on a single RTX 3090 unquantized. To try 8-bit mode, set load_in_8bit=True in chat.py

How to distill the weights

1. Put LLaMA weights into original/ folder, such that 7B version would be at original/7B

2. Download point-alpaca diffs into encrypted/ folder:

wget -P encrypted/ -i filelist.txt

3. Run the following command to decrypt:

for f in "encrypted"/*; do if [ -f "$f" ]; then python3 decrypt.py "$f" "original/7B/consolidated.00.pth" "result/"; fi; done

Windows users can use the equivalent powershell command:

Get-ChildItem "encrypted" | ForEach-Object {
    if($_.Attributes -eq 'Archive') {
        python3 decrypt.py $_.FullName "original/7B/consolidated.00.pth" "result/"
    }
}

You will have finetuned weights in the result/ folder.

Now that you have them, you can delete the files in encrypted/ folder.

How to chat with the model

Other people will probably build better UIs, but for now, try running python3 chat.py

But before that, install requirements via pip3 install -r requirements.txt (We really recommend installing it in a separate environment, for example, via conda)

Questions? Suggestions?

Find us in our Telegram chat: https://t.me/pointnetworkchat

Why are weights "encrypted"?

We are not allowed to publish weights for LLaMA, of course, even finetuned, but there is no problem publishing the difference, a patch that we suggest to apply to the files. The encryption is a simple XOR between files (not very secure - not recommended for other applications!), ensuring that only the people that have access to the original weights (from completely legal sources, of course) can transform them into finetuned weights.

What are the checksums so I can check if something is wrong?

$ md5sum encrypted/*
4b8622230b59b3f3bcad791c8c1bae51  encrypted/added_tokens.json.75e3ca5df2973756aa612cb17246ef6020a68ff8d94671508987d373642f7a36.enc
876376085d79041818bb7a41bced7819  encrypted/config.json.caf9cac32580e31af8254f66c5a070741d70b15a651721748189180325b7d5a8.enc
44b1feec4c0d1b7c87da24b81c8b8b9e  encrypted/generation_config.json.c5c8961ed243834883fb4e45e8850d3873d6100fde97817f59d275a90eba269d.enc
d127aabb6ad5375bfa97c6ac529c166d  encrypted/pytorch_model-00001-of-00003.bin.90d2ab95a32aeb9362814d8b86db2af5454baab8ea3aa8230c271d6962abb9db.enc
e4b12501e99cf6a30a244af20f5c20ec  encrypted/pytorch_model-00002-of-00003.bin.f3c10a4f5c8beafc6667d34557b64ba479e4dde6ef10672287857b329b7e3229.enc
d212294c06feeb0f14672b68417dbc9e  encrypted/pytorch_model-00003-of-00003.bin.72bf4c96aa6b0c7b56b0336791960da9c75de324ea1131ea4bfc20fde41115c8.enc
e813854dede95a03e5f5b459c7fb32b2  encrypted/pytorch_model.bin.index.json.07ca8edea996b6c3274395fdb2b6c9108f2ffdd610ae55e35c126c21a9d535b1.enc
62503bbf4e91f2b50bf9834757d555d3  encrypted/special_tokens_map.json.4ad09c72922c015ba04f09eabebe38fb34ecb721ca712922c62038eaf2d0bc61.enc
39ec1b33fbf9a0934a8ae0f9a24c7163  encrypted/tokenizer.model.9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347.enc
2c34a03919b6b2b299ad6f77713d0ba0  encrypted/tokenizer_config.json.a5f5efb2240276709a923b1404e08d93cc896fd1bd31fbe173e1e2789ea210ef.enc
560ecf526666cbd485b81f0f16bb9972  encrypted/trainer_state.json.43964ae247e74f4055fe1cf99a7a16efc3114402a1cd918b3cd9e2ebf2858ca9.enc
fe8b25ba7c8dd66d57ce1d3d60f13abd  encrypted/training_args.bin.02f8c3ba14e3c48c05f76880975d7385c878b0e5a0863e352c82f331150d2bd4.enc

$ md5sum original/7B/consolidated.00.pth
6efc8dab194ab59e49cd24be5574d85e  original/7B/consolidated.00.pth

$ md5sum result/*
880c59f7618832454595e9820960c360  result/added_tokens.json
d39ed682be60de38e12c5d1974c45620  result/config.json
5300908d1f82b0bc7a4bc79ea00dad66  result/generation_config.json
5d17f8837f9f15538acd65b7d37add2c  result/pytorch_model-00001-of-00003.bin
834b0748527482d60236bc1ec0c71750  result/pytorch_model-00002-of-00003.bin
03dda8d1057b06632fecf399020353b4  result/pytorch_model-00003-of-00003.bin
82559775d42e04199b5a8be8df974b36  result/pytorch_model.bin.index.json
40df8792c753f0d3f5786829efdd2954  result/special_tokens_map.json
eeec4125e9c7560836b4873b6f8e3025  result/tokenizer.model
f2da7d9c67a3b7d2e60a17c540055b85  result/tokenizer_config.json
883795093c1f18baa9b111880b800bf1  result/trainer_state.json
f07e553d22ebe37908bc996953f1bb11  result/training_args.bin

What about larger models?

13B is coming for sure, larger versions - maybe. Consider supporting us if you want it done faster. :)