Update README.md
Browse files
README.md
CHANGED
|
@@ -950,11 +950,6 @@ custom code or config file is required.
|
|
| 950 |
VQ-VAE's downsampling factor).
|
| 951 |
- No instruction-tuning or RLHF; this is a base, unaligned model trained
|
| 952 |
purely on the next-token objective across the unified token stream.
|
| 953 |
-
- The two undocumented architectural assumptions described above (attention
|
| 954 |
-
head count, decoder output activation) were inferred from common
|
| 955 |
-
conventions, not recovered from the checkpoint itself. If you trained
|
| 956 |
-
this model and know the real values, please open a discussion or a PR
|
| 957 |
-
with the corrected `run_supra_a2a.py` config.
|
| 958 |
|
| 959 |
## About SupraLabs
|
| 960 |
|
|
|
|
| 950 |
VQ-VAE's downsampling factor).
|
| 951 |
- No instruction-tuning or RLHF; this is a base, unaligned model trained
|
| 952 |
purely on the next-token objective across the unified token stream.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 953 |
|
| 954 |
## About SupraLabs
|
| 955 |
|