rwkv-4-pile-14b / README.md
BlinkDL's picture
Update README.md
3307c7e
metadata
language:
  - en
tags:
  - pytorch
  - text-generation
  - causal-lm
  - rwkv
license: apache-2.0
datasets:
  - the_pile

RWKV-4 14B

[UPDATE: Try RWKV-4-World (https://huggingface.co/BlinkDL/rwkv-4-world) for generation & chat & code in 100+ world languages, with great English zero-shot & in-context learning ability too.]

Model Description

RWKV-4 14B is a L40-D5120 causal language model trained on the Pile. See https://github.com/BlinkDL/RWKV-LM for details.

args.n_layer = 40 args.n_embd = 5120

Use https://github.com/BlinkDL/ChatRWKV to run it.

RWKV-4-Pile-14B-2023xxxx-ctx8192-testxxx.pth : Fine-tuned to ctx_len 8192.

  • The best general model.

################################

"Raven": RWKV alpaca+vicuna-style model: https://huggingface.co/BlinkDL/rwkv-4-raven (highly recommended)

It is a strong chat model too. You can use +i for "Alpaca Instruct" in latest ChatRWKV v2. Examples:

+i Explain the following metaphor: "Life is like cats". 
+i write a python function to read data from an excel file.

################################

RWKV-4-Pile-14B-20230213-8019.pth : Trained on the Pile for 331B tokens

  • Pile loss 1.7579 (ctx_len 1024)
  • LAMBADA ppl 3.81, acc 71.05%
  • PIQA acc 77.42%
  • SC2016 acc 75.57%
  • Hellaswag acc_norm 70.24%
  • WinoGrande acc 62.98%