minbert-sts-nli-meanpool

A bert-base-uncased encoder fine-tuned in two stages for Semantic Textual Similarity (STS-B style regression, labels in [0, 5]):

1. NLI pretraining. The bi-encoder is first trained as a Siamese sentence encoder on the combined SNLI + MultiNLI entailment task (942,069 train pairs), giving it a head start at telling semantically related sentences apart before ever seeing an STS label.
2. STS fine-tuning. The NLI-pretrained encoder is then fine-tuned on the STS training set: all layers unfrozen, lr 5e-6, batch size 32, 100-step linear warmup, weight decay 0.05, dropout 0.2, up to 4 epochs with early-stopping patience 2.

Scoring. There is no learned similarity head. A sentence is represented by masked mean pooling over last_hidden_state (including the [CLS]/[SEP] positions), and the similarity score is the cosine of the two sentence vectors, rescaled from [-1, 1] to [0, 5]:

score = (cosine_similarity(mean_pool(s1), mean_pool(s2)) + 1) * 2.5

Mean pooling was chosen over the [CLS]/pooler-output representation because it outperformed CLS at every layer in a pooling-strategy sweep on this project's STS dev set.

Results (STS-B dev set)

Seed	Pearson	Spearman
11711	0.7432	0.7457
2026	0.7436	0.7460
42	0.7388	0.7411
mean (n=3)	0.7419	0.7443

The seed=11711 checkpoint published here (Pearson 0.7432, Spearman 0.7457) is the one submitted for this project; report the 3-seed mean above, not the single-seed number, as the figure the method should be judged on -- std over 3 seeds is a rough sense of spread, not a precise estimate.

Usage

This checkpoint stores only the BERT encoder (bert.*); there is no separate classification head to load. To reproduce the pooling/scoring exactly, use this project's bert.BertModel and multitask_classifier.masked_mean_pool (https://github.com/nirish2407/dnlp), or replicate the same masked mean pooling + cosine recipe with any BERT loader.

Training details

See config.json in this repository for the full recorded training recipe (learning rate, epochs, dropout, warmup, weight decay, and the NLI checkpoint this model was initialised from), and the STS section of the project README (https://github.com/nirish2407/dnlp) for the full experimental writeup, including the ablations that motivated each choice (NLI-pretraining control run, pooling-strategy sweep, layer-wise readout analysis, and a residualized probe checking the gains are not just lexical overlap).