107 lines
4.4 KiB
Markdown
107 lines
4.4 KiB
Markdown
---
|
|
license: mit
|
|
language:
|
|
- fr
|
|
tags:
|
|
- babylm
|
|
- babylm-2026
|
|
- gpt2
|
|
- causal-lm
|
|
- french
|
|
- cross-linguistic
|
|
- low-resource
|
|
- right-tool-right-job
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# BabyLM 2026 Strict, French (92M words)
|
|
|
|
A 125M-parameter GPT-2 trained from scratch on 92,469,402 words of French text. Submitted to the BabyLM 2026 Strict track and the primary checkpoint reported in *Right Tool, Right Job: Why Training Language Matters More Than Training Data* (Wasserman & Beauchemin, BabyLM 2026 / ACL Rolling Review submission).
|
|
|
|
## Headline result
|
|
|
|
QFrBLiMP (Quebec French native minimal-pair benchmark, 1761 pairs): **85.97% overall**.
|
|
|
|
| Subset | Pairs | Accuracy |
|
|
|---|---:|---:|
|
|
| Anglicism | 267 | 80.15% |
|
|
| Morphology | 716 | 85.47% |
|
|
| Semantic | 398 | 87.19% |
|
|
| Syntax | 380 | 89.74% |
|
|
| **Overall** | **1761** | **85.97%** |
|
|
|
|
QFrCoLA (Quebec French acceptability classification, fine-tuned with LoRA rank 16): test accuracy ~72%, MCC ~0.24 (epoch 3 of fine-tune).
|
|
|
|
## Argument supported by this model
|
|
|
|
The companion paper develops the cross-linguistic argument that training-language morphological richness, not neural architecture or pretraining scale, is the load-bearing variable for grammar acquisition. This checkpoint is the child-scale (under 100M words) French anchor; the broader argument is also supported by the *Scaling Hypothesis Is Language-Contingent* and *English Considered Harmful* deposits cited below, which test the same claim at different scales and with different ablations.
|
|
|
|
## Model details
|
|
|
|
- Architecture: GPT-2 decoder-only, causal LM (`GPT2LMHeadModel`)
|
|
- Parameters: ~125M
|
|
- Layers: 12
|
|
- Attention heads: 12
|
|
- Hidden size: 768
|
|
- Max sequence length: 512
|
|
- Vocabulary: 50,000 BPE, French Wikipedia source
|
|
- Precision: float32
|
|
|
|
## Training data
|
|
|
|
- 92,469,402 words of French (under the BabyLM 2026 Strict 100M-word cap)
|
|
- Custom corpus assembled from CHILDES French subsets and the Orléans corpus as a developmental base, with lemma-frequency oversampling guided by a Haitian Creole vocabulary oracle (high-frequency, high-composability lemmas surviving pidginization)
|
|
- Training data is 100% morphologically rich French; Haitian Creole sentences are not mixed in
|
|
- See *Right Tool, Right Job* §3 for full corpus curation methodology
|
|
|
|
## Training procedure
|
|
|
|
- Peak learning rate: 1.0e-4
|
|
- LR schedule: cosine decay to ~1.9e-7
|
|
- Epoch: 3 (of a 5-epoch trajectory; epoch 3 is the grammatical-competence peak reported in §4.2 of the paper)
|
|
- Tokens/sec: ~94,000 (CUDA)
|
|
- Approximate GPU hours through epoch 3: ~3
|
|
- Final training loss: 3.19, perplexity 24.4
|
|
|
|
## Intended use
|
|
|
|
Suitable for:
|
|
- Replicating *Right Tool, Right Job* results
|
|
- Cross-linguistic emergence research
|
|
- Quebec French native-benchmark development
|
|
- Studies of morphological redundancy and training-data efficiency at child scale
|
|
|
|
Not suitable for:
|
|
- General-purpose French text generation at production quality (corpus is developmental, not web-scale)
|
|
- Any English-language task (the model has zero English training exposure)
|
|
|
|
## Limitations
|
|
|
|
- French-only training; zero exposure to English or other non-French data
|
|
- Child-scale corpus (92M words) is far below typical web-scale pretraining
|
|
- BPE tokenizer trained on French Wikipedia, which differs in register from the CHILDES / Orléans developmental sources
|
|
- LoRA fine-tuning was used in downstream evaluation grids (see *Right Tool, Right Job* §5)
|
|
|
|
## Citation
|
|
|
|
```bibtex
|
|
@inproceedings{wasserman_beauchemin_2026_right_tool,
|
|
title = {Right Tool, Right Job: Why Training Language Matters More Than Training Data},
|
|
author = {Wasserman, Adam Z. and Beauchemin, David},
|
|
booktitle = {BabyLM 2026 Workshop / ACL Rolling Review submission},
|
|
year = {2026}
|
|
}
|
|
```
|
|
|
|
Companion deposits supporting the broader cross-linguistic argument:
|
|
|
|
- Wasserman, Adam Z. (2026). *The Scaling Hypothesis Is Language-Contingent.* Zenodo DOI [10.5281/zenodo.19423151](https://doi.org/10.5281/zenodo.19423151).
|
|
- Wasserman, Adam Z. (2026). *English Considered Harmful: How Morphological Poverty Pollutes Language Model Training.* Zenodo DOI [10.5281/zenodo.19443357](https://doi.org/10.5281/zenodo.19443357).
|
|
|
|
Pre-registrations on OSF: [SJ48B](https://osf.io/sj48b) (Language-Only Hypothesis), [PCX2D](https://osf.io/pcx2d) (morphological complexity gradient).
|
|
|
|
## Acknowledgments
|
|
|
|
The QFrBLiMP and QFrCoLA evaluation benchmarks are by David Beauchemin and collaborators (Université Laval, Institut Intelligence et Données).
|