diff options
| author | Matthew Li <156706407+mattqlf@users.noreply.github.com> | 2026-03-19 17:00:42 -0400 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2026-03-19 14:00:42 -0700 |
| commit | 3a6fec7941f3cc4187ac208496b6842f3563e18c (patch) | |
| tree | 928fdd23b019ea48234189fd45bb729adba8a991 /records | |
| parent | d84a3e819100504d96879e1e36d022efa5cbb81b (diff) | |
Fix: score final partial window in sliding window eval (#124)
The window_starts filter dropped windows shorter than stride,
silently skipping up to (stride-1) tokens at the end of the
validation set. Now includes all windows with >= 1 scoreable
token, and clamps the score start for short final windows.
Diffstat (limited to 'records')
| -rw-r--r-- | records/track_10min_16mb/2026-03-19_SlidingWindowEval/train_gpt.py | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/records/track_10min_16mb/2026-03-19_SlidingWindowEval/train_gpt.py b/records/track_10min_16mb/2026-03-19_SlidingWindowEval/train_gpt.py index 8555de6..6a8fd84 100644 --- a/records/track_10min_16mb/2026-03-19_SlidingWindowEval/train_gpt.py +++ b/records/track_10min_16mb/2026-03-19_SlidingWindowEval/train_gpt.py @@ -856,9 +856,9 @@ def eval_val_sliding( seq_len = args.train_seq_len total_tokens = val_tokens.numel() - 1 - # Build windows; skip any too short to score a full stride + # Build windows; include final partial window if it has at least 1 token window_starts = [ws for ws in range(0, total_tokens, stride) - if min(ws + seq_len, total_tokens) - ws >= stride] + if min(ws + seq_len, total_tokens) - ws >= 1] total_windows = len(window_starts) # Distribute across ranks @@ -899,7 +899,7 @@ def eval_val_sliding( for i, ws in enumerate(batch_ws): wlen = wlens[i] - s = 0 if ws == 0 else wlen - stride + s = 0 if ws == 0 else max(wlen - stride, 0) scored_nll = nll[i, s:wlen].to(torch.float64) loss_sum += scored_nll.sum() token_count += float(wlen - s) |