From 8712d595f1ff735a8634a2105b0e46c29c16341a Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Thu, 8 Oct 2026 06:43:31 +0800 Subject: [PATCH] fix: inv vocab scope in thai tiny decode (resumes from saved student) --- train_thai_tiny_hf.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/train_thai_tiny_hf.py b/train_thai_tiny_hf.py index 57470a0..446ffd5 100644 --- a/train_thai_tiny_hf.py +++ b/train_thai_tiny_hf.py @@ -239,6 +239,7 @@ def collate(b): # ---- greedy decode kaikki-test, corpus PER ---- model.eval() + inv = {v: k for k, v in vocab.items()} preds = [] with torch.no_grad(): for i in range(0, len(test), 64): @@ -266,7 +267,6 @@ def collate(b): for row in ys[:, 1:].tolist(): toks = [c for c in row if c not in (pad, eos)] preds.append("".join(inv[t] for t in toks)) - inv = {v: k for k, v in vocab.items()} per = corpus_per(list(zip(preds, [t for _, t in test]))) print(f"[gate] val PER (greedy, corpus) = {per:.4f}% gate<={GATE_PER}%", flush=True)