jamesdavie/coca-matching

Coca experiments: Matching

Overview

This is an experimental Coca codebase for Matching. It keeps the large setup intentionally manageable so architecture changes can be inspected before a full training run.

Expect rough edges — this is a working sketch, not a polished release.

Contact

Questions or corrections: open a discussion in this repository.

Limitations

The initialization checkpoint has not been trained or audited for robustness, fairness, or domain transfer. The implementation should be treated as an experimental starting point. Results from a future trained checkpoint must be documented separately from the defaults shipped here.

Repository Status

One Python module contains the model and a runnable example. Generated configuration lives in config.json; the recipe in training_args.json. The checkpoint is an initialization artifact and should not be read as a trained model. Scores are deliberately absent.

Architecture

Item	Value
Architecture	Coca
Scale	large
Attention	standard
Fusion	tucker
Activation	gelu tanh
Normalization	groupnorm

Default Experiment Recipe

The generated recipe uses adam with a exponential schedule.

These values exist so the script runs end-to-end. A meaningful experiment needs its own tuning budget and at least three seeds.

Evaluation Guidance

A first evaluation could use a paired validation set. Report the metric across at least three seeds and include a matched-capacity baseline; keep logs and library versions with any published number.

Quick Check

Run python finetune.py --help for the built-in smoke test. Because this is a custom implementation, generic automatic loading APIs require an explicit adapter before use.

Files

• finetune.py — primary artifact
• README.md — this documentation
• config.json — architecture configuration
• training_args.json — default experiment settings
• model.safetensors — initialization checkpoint
• USE_POLICY.md — use and redistribution notes

License

Distributed under the unknown license. External datasets keep their own terms — check them separately.