---
language:
- en
license: mit
library_name: transformers
pipeline_tag: text-generation
tags:
- babylm
- babylm-2026
- strict
- custom_code
---

# BabyLM Challenge 2026 submission | RecGPT-100M

RecGPT-100M is a 124.03M-parameter recursive causal language model trained for the BabyLM 2026 Strict track. It was trained for 10 epochs on a custom 100M-word English corpus using a 32,768-token BPE vocabulary.

The model applies a shared Transformer block recursively for 24 iterations. Its hidden size is 1,408, embedding size is 768, and feed-forward intermediate size is 22,528. Training used Aurora for the recursive block and AdamW for the embedding-related parameters, with a token batch size of 32,768 and sequence length 512.

## Usage

This repository contains custom Transformers code, so loading requires `trust_remote_code=True`:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Serdar404/RecGPT-100M-Fixed"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
```

The model is intended for scoring text as a causal language model. KV-cache generation is not currently implemented.

## BabyLM 2026 evaluation

Final-checkpoint results before leaderboard submission:

| Evaluation | Score |
|---|---:|
| BLiMP | 80.06 |
| BLiMP Supplement | 69.28 |
| EWoK | 59.05 |
| Entity Tracking | 19.50 |
| COMPS | 60.55 |
| GlobalPIQA | 43.15 |
| (Super)GLUE | 71.84 |

Intermediate Strict checkpoints are published as Hub revisions named `chck_1M` through `chck_1000M` using the official BabyLM checkpoint schedule.

## Resources

- Training code: https://github.com/serdardoesml/bblm26-recgpt
- Dataset construction: https://github.com/serdardoesml/bblm26-dataset
- Evaluation fork: https://github.com/serdardoesml/babylm-eval

## Limitations

This is a small research model trained under the BabyLM data constraint. It is not intended for production deployment, factual question answering, or safety-critical use. Its outputs may contain inaccuracies or undesirable content inherited from its training data.