T5-small — Cross-Attention-Only fine-tune on SQuAD
Encoder–decoder QA model fine-tuned from google-t5/t5-small (60.5M) by training only the
decoder cross-attention (EncDecAttention) blocks plus the decoder final layer norm —
~6.3M trainable params (10.4%). The encoder, decoder self-attention, and feed-forward
weights are frozen.
TL;DR: a 60M-param encoder–decoder, tuning just 6.3M of those weights, reaches F1 0.7075 / EM 0.5293 on SQuAD — outperforming a fully fine-tuned GPT-2-large (774M, F1 0.5041) that is 13× larger. The strongest evidence in this study that architecture beats scale for context-grounded tasks.
Results (SQuAD validation)
| Model | Architecture | Trainable / Total | EM | Token F1 |
|---|---|---|---|---|
| GPT-2-large | decoder-only | 774M / 774M | 0.3516 | 0.5041 |
| T5-small XA-only (this model) | enc-dec | 6.3M / 60.5M | 0.5293 | 0.7075 |
| T5-large XA-only | enc-dec | 100.7M / 737M | 0.6406 | 0.8128 |
| T5-large LoRA r=8 | enc-dec | 2.4M / 740M | 0.6445 | 0.8152 |
Validation loss 0.4208, perplexity 1.52. Trained on only 3,000 SQuAD examples.
Eval note: 256-example SQuAD-validation generation subset, 4-beam search.
How to use
Trained with the input prefix answer question: prepended to a
question: ... context: ... source string — match it exactly:
from transformers import T5ForConditionalGeneration, AutoTokenizer
repo = "medelharchaoui/t5-small-xa-only-squad"
tok = AutoTokenizer.from_pretrained(repo)
model = T5ForConditionalGeneration.from_pretrained(repo)
question = "What culture do 'bairn' and 'hyem' originate from?"
context = ("'bairn' and 'hyem' are geordie words with origins in scandinavia; barn and hjem "
"are the corresponding modern norwegian and danish words.")
text = f"answer question: question: {question} context: {context}"
ids = tok(text, return_tensors="pt", truncation=True, max_length=384).input_ids
print(tok.decode(model.generate(ids, num_beams=4, max_new_tokens=16)[0], skip_special_tokens=True))
Training
| Setting | Value |
|---|---|
| Base model | google-t5/t5-small (60.5M) |
| Trainable params | *.EncDecAttention.* + decoder.final_layer_norm (~6.3M) |
| Dataset | rajpurkar/squad, 3,000 train examples |
| Precision | bf16 |
| Optimizer steps | 1,500 (batch 4 × grad-accum 8 = eff. batch 32) |
| LR / warmup | 2e-4, 150 warmup, weight decay 0.01 |
| Source / target max len | 384 / 32 |
| Seed | 37 |
| Hardware | 1× NVIDIA RTX 3060 (12 GB), local |
Limitations
- English SQuAD-style extractive QA only; short answer spans grounded in the supplied context.
- Small model trained on only 3,000 examples — strong for its size on SQuAD, but not a general-purpose QA system.
- Evaluated on a held-out validation subset, not the official SQuAD test server.
Citation
Part of an encoder–decoder vs decoder-only paradigm study (OptimiAI, 2026).