tesda-e5-small-v4-colab-1

A retraining run of the TESDA-TRACK matching model, made with the v4 training data in Google Colab (round v4-colab-1). It is a fine-tune of intfloat/multilingual-e5-small that matches learner goals to TESDA qualifications. The training data covers English, Filipino and Taglish, with synthetic examples in Cebuano, Ilocano, Hiligaynon and Kapampangan.

Status: not yet evaluated. This run hasn't been measured on the dev, sealed or v1-test sets, and it has no calibration.json, so the TESDA-TRACK API scores it with the settings' default parameters. Its regional-language examples are synthetic drafts that fluent speakers haven't reviewed. For serving, use tesda-e5-small-v3, the recommended model, or the evaluated tesda-e5-small-v4.

The repo contains the ONNX model only, with no PyTorch weights.

Set up in TESDA-TRACK

From the repository root:

hf download Nyanta19/tesda-e5-small-v4-colab-1 --local-dir backend/models/tesda-e5-small-v4-colab-1

Before serving it, measure it as described in section 4 of backend/training/V4.md. To try it, set both lines in the root .env, run python -m tesda_track.embeddings sync from backend/, and restart the API:

EMBEDDING_MODEL=tesda-track/tesda-e5-small-v4-colab-1
EMBEDDING_MODEL_PATH=backend/models/tesda-e5-small-v4-colab-1

Stored embeddings are keyed by EMBEDDING_MODEL, so keep the name exactly as given.

Files

File	Purpose
onnx/model.onnx	The model (470 MB, fp32)
tokenizer.json, tokenizer_config.json, config.json	Tokenizer and model config
training.json, run_config.json	Training run record and settings
coverage.json	Training-data counts per language, example kind and qualification
checksums.json, bundle_sha256.txt	SHA-256 of the files produced by the training run, and of the training bundle

Prefix goals with query: and qualification text with passage: . Embeddings have 384 dimensions, mean-pooled and L2-normalised; compare them with cosine similarity.

Training

From training.json:

• 45,674 pairs (43,395 train, 2,279 held out), including 25,199 synthetic and 2,166 augmented pairs.
• 3 epochs, 2,037 steps, batch size 64 (mini-batch 8), learning rate 2e-5, seed 42, on a CUDA GPU in 56 minutes.
• Held-out loss fell from 2.98 to 0.111. ONNX export parity cosine: 0.9999993.
• Trained 22 October 2026.

Training loss and export parity don't establish matching accuracy in any language.