RAFT shenanigans

2026-03-23 08:52:42 +01:00 · 2026-02-21 23:47:12 +01:00
parent 49c622db08
commit 61edb35f70
14 changed files with 2943 additions and 6 deletions
--- a/raft/train_mistral_raft.py
+++ b/raft/train_mistral_raft.py
@@ -73,7 +73,6 @@ def main():
        max_length=args.max_seq_len,
        bf16=torch.cuda.is_available(),
        fp16=not torch.cuda.is_available(),
-        assistant_only_loss=True,  # only learn from assistant turns in messages
        report_to=[],
    )