bert-base-uncased-snli-5k
A deliberately weak, non-robust NLI baseline: bert-base-uncased fine-tuned on only 5,000 SNLI training pairs.
Training
- Data: 5,000 random gold-labelled pairs from the SNLI training set (seed 0).
- 3 epochs, learning rate 2e-5 (linear schedule, 10% warmup), batch size 32, AdamW, max length 128.
Labels
0 = entailment, 1 = neutral, 2 = contradiction (SNLI order), named in the config.
Input: tokenizer(premise, hypothesis).
Accuracy
| Dataset | Accuracy |
|---|---|
| SNLI validation | 0.782 |
| MNLI validation matched | 0.530 |
| MNLI validation mismatched | 0.551 |