recurrent-transformer-gpt2-sft

Recurrent Transformer GPT-2

Small Transformer-XL style recurrent language model trained with a GPT-2 tokenizer.

Model

• Architecture: decoder-only recurrent transformer with detached KV memory
• Tokenizer: GPT-2 tokenizer (vocab_size=50257)
• Hidden size: 512
• Layers: 6
• Heads: 8
• Segment length: 128
• Memory length: 256
• Effective attention window: 384 tokens

Training

Pretraining used local parquet text shards copied from the neighboring project, then tokenized with the GPT-2 tokenizer.

The preserved pretrained weights are:

• base.pt: best pretraining checkpoint
• latest_pretrain.pt: latest pretraining checkpoint

The full local checkpoint history is intentionally not included in this model upload because it is large. It is backed up locally under:

backups/pretrain_2026-06-17/

Fine-Tuning Data

The selected SFT dataset is databricks/databricks-dolly-15k, downloaded locally as:

data/sft/raw/databricks-dolly-15k.jsonl

It is suitable for this small model because it is compact, human-written, and single-turn instruction/QA oriented.

Intended Use

Research and experimentation with small recurrent transformer pretraining and SFT.