Nyanta19/tesda-e5-small-v6-colab

TESDA E5 Small v6 (Colab)

ONNX embedding model for matching learner goals to TESDA qualifications. Continued training from Nyanta19/tesda-e5-small-v4 using the v6 multi-positive objective and unrelated-goal rejection examples.

This repository contains the exact serving files from tesda-e5-small-v6-colab.zip. File hashes are recorded in checksums.json. The ONNX encoder outputs 384-dimensional token embeddings; apply attention-mask mean pooling and L2 normalization, using the E5 query: and passage: input prefixes.

Trained on a Tesla T4 for 1,132 steps (one epoch, learning rate 1e-5). See training.json and run_config.json for training settings and provenance. Training held-out loss decreased from 0.3165 to 0.2810. These training diagnostics are not application evaluation results.

This Colab artifact contains no calibration.json. Calibrate and evaluate this specific run before using it for application scoring or rejection.

The export contains ONNX weights rather than PyTorch or safetensors weights.