|
--- |
|
language: |
|
- en |
|
tags: |
|
- pytorch |
|
- text-generation |
|
- causal-lm |
|
- rwkv |
|
license: apache-2.0 |
|
datasets: |
|
- the_pile |
|
|
|
--- |
|
|
|
# RWKV-4 3B |
|
|
|
# Use RWKV-4 models (NOT RWKV-4a, NOT RWKV-4b) unless you know what you are doing. |
|
# Use RWKV-4 models (NOT RWKV-4a, NOT RWKV-4b) unless you know what you are doing. |
|
# Use RWKV-4 models (NOT RWKV-4a, NOT RWKV-4b) unless you know what you are doing. |
|
|
|
## Model Description |
|
|
|
RWKV-4 3B is a L32-D2560 causal language model trained on the Pile. See https://github.com/BlinkDL/RWKV-LM for details. |
|
|
|
Use https://github.com/BlinkDL/ChatRWKV to run it. |
|
|
|
RWKV-4-Pile-3B-20221110-ctx4096.pth (RECOMMENDED) : Fine-tuned to ctx_len 4096. |
|
* LAMBADA ppl 5.25, acc 63.96% |
|
* PIQA acc 74.16% |
|
* SC2016 acc 70.71% |
|
* Hellaswag acc_norm 59.89% |
|
* ctx_len = 4096 n_layer = 32 n_embd = 2560 |
|
|
|
RWKV-4-Pile-3B-20221008-8023.pth : Trained on the Pile for 331B tokens. |
|
* Pile loss 1.9469 |
|
* LAMBADA ppl 5.24, acc 63.94% |
|
* PIQA acc 73.72% |
|
* SC2016 acc 70.28% |
|
* Hellaswag acc_norm 59.63% |
|
* ctx_len = 1024 n_layer = 32 n_embd = 2560 |
|
|
|
### Instruct-test models: only useful if you construct your prompt following dataset templates |
|
|
|
Note I am using "Q: instruct\n\nA: result" prompt for all instructs. |
|
|
|
RWKV-4-Pile-3B-Instruct-test1 |
|
instruct-tuned on https://huggingface.co/datasets/bigscience/xP3all/viewer/en/train |
|
|
|
RWKV-4-Pile-3B-Instruct-test2 |
|
instruct-tuned on https://huggingface.co/datasets/Muennighoff/flan & NIv2 |
|
|
|
### Chinese models |
|
|
|
RWKV-4-Pile-3B-EngChn-testNovel-xxx for writing Chinese novels (trained on 200G Chinese novels.) |
|
|
|
RWKV-4-Pile-3B-EngChn-testxxx for Chinese Q&A (trained on 10G Chinese text. only for testing purposes.) |
|
|
|
## Note: 4 / 4a / 4b models ARE NOT compatible. Use RWKV-4 unless you know what you are doing. |
|
|
