| name | tenacious-training-prep |
| description | Convert Tenacious-Bench artifacts into high-quality training inputs for Path A, Path B, or Path C. Use when an agent is formatting chat pairs, preference pairs, step-level labels, quality filters, or path-specific training partitions for LoRA, judge training, or process reward modeling in this project. |
Tenacious Training Prep
Overview
Good training runs start with disciplined training data, not optimistic hyperparameters.
Read references/path-formats.md before building or filtering training data.
Workflow
1. Choose the path from evidence
Do not prepare all three formats just in case. Build the format that matches the chosen intervention.
2. Filter before formatting
Prefer fewer cleaner examples over many noisy ones. Remove:
- broken records
- ambiguous labels
- low-confidence rewrites
- examples that conflict with the current rubric
- contamination risks against dev or held-out
3. Format for the selected path
- Path A: clean input/output chat pairs
- Path B: chosen/rejected preference pairs with leakage-aware generation
- Path C: stepwise trajectories with intermediate correctness labels
Keep formatting explicit and reproducible.
4. Preserve provenance
Carry forward:
- source task IDs
- rewrite origin
- quality filter reason
- path designation
- split origin
5. Treat flat ablations as a data warning
If a training run does not lift, inspect the training data before touching compute or prompts.
Output Shape
Prefer machine-friendly exports and a small methodology note that explains:
- what was included
- what was filtered out
- which model family produced any rewrites
- how preference leakage was avoided when relevant
Practical Defaults
- Quality beats quantity at this project scale.
- Keep path-specific formats simple and inspectable.
- Reuse the evaluator to filter training examples whenever possible.
