Fixes
- CI test fix: Strip ANSI escape codes in
test_tensorboard_in_train_help— Rich splits--tensorboardacross escape sequences on Python 3.11, causing string assertion to fail - TensorBoard import check: Use
import tensorboarddirectly instead oftorch.utils.tensorboardshim for accurate availability detection - Streaming output:
soup infernow writes results to disk as they're generated (crash-safe for large prompt files) - Accurate token count:
_generatereturns token count from tensor shape instead of re-encoding decoded text - Security:
max_tokenscapped at 16384,trust_remote_codewarning shown before model load - Better tests: Replaced shallow inline-logic tests with real trainer integration tests that verify
report_to='tensorboard'is accepted by SFT/DPO/GRPO/KTO wrappers
Install / Upgrade
pip install --upgrade soup-cli