Known Issues

Known Issues and Implementation Gaps

Severity	Area	Finding / next step
Blocking	DPO_SFT.py	Syntax error at line 44 (unterminated string); indentation and other issues follow. Keep labeled WIP until repaired.
High	Preference pipeline	No validated Phi-2 training, checkpoint reload, DPO reference-policy test, or preference evaluation.
High	TRL / transformers compatibility	Verify actual SFTTrainer/DPOTrainer signatures against installed versions; draft uses API patterns that may have changed.
High	Prompt formatting	Hardcoded `<
Medium	Source comments	The original README and requirements.txt incorrectly refer to DPO_PPO.py; actual filename is DPO_SFT.py. The published README corrects this, while upstream originals are preserved.
Medium	Dataset	30 classifier rows with only six validation cases; 6 SFT examples and 5 preference pairs.
Medium	Training compatibility	DistilBERT helper uses version-dependent signatures; runtime validation still needed.
Medium	Missing continuous integration	There are no bundled end-to-end tests or pinned lockfiles.
Info	PPO / ORPO	PPO is absent; ORPO appears only as a printed string and is not run.
Info	Model outputs	No pre-trained artifact or measured benchmark is supplied in ZIP.

See Validation-Report and Roadmap.