Add unsupervised coefficient tuning to model-merging skill

[ADu2021]

Summary

Extends the existing 19-emerging-techniques/model-merging skill with an advanced, label-free method for selecting merge coefficients automatically, based on arXiv:2503.23733.

Contents

• SKILL.md (+17 lines): Added concise description and pseudocode under Best Practices → Weight Selection, with a link to the new reference file
• references/coefficient-tuning.md (411 lines, ~12KB): Full technical reference for the generation consistency method

What It Covers

The new reference documents an unsupervised alternative to manual or grid-search coefficient tuning:

1. Merge models across N candidate coefficients
2. Run inference on a small unlabeled data subset (~50–200 prompts)
3. For each candidate α, measure similarity of its outputs to those of its neighbors (α ± δ)
4. Select the coefficient with the highest generation consistency score

Source

• Paper: arXiv:2503.23733 (CVPR 2025 paper)

Documentation Size

• New reference file: 411 lines / ~12KB
• Total references directory: 52KB

Key Features Included

• Full 5-step algorithm with copy-paste Python implementation
• Three similarity metric options: token overlap, ROUGE-L, BERTScore (with trade-off table)
• Guidance for applying the method to SLERP, Task Arithmetic, and TIES
• Coordinate-wise search for multi-coefficient (3+ model) scenarios
• End-to-end unsupervised_coefficient_search() pipeline function
• Practical tips on dataset size, step size δ, boundary candidates, and flat-curve diagnostics