初始化项目，由ModelHub XC社区提供模型

Model: openhonest/babylm-2026-fr-92m
Source: Original Platform

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,106 @@
+---
+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).

config.json

@@ -0,0 +1,34 @@
+{
+  "activation_function": "gelu_new",
+  "add_cross_attention": false,
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 0,
+  "dtype": "float32",
+  "embd_pdrop": 0.1,
+  "eos_token_id": 0,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 512,
+  "pad_token_id": 1,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.5.3",
+  "use_cache": true,
+  "vocab_size": 50000
+}

generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 0,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "transformers_version": "5.5.3",
+  "use_cache": true
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d4c342119eef6243b24b4f5224334909d4896a113ec22739075dc4f5c384638
+size 495411840

tokenizer_config.json

@@ -0,0 +1,8 @@
+{
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "model_max_length": 512,
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "unk_token": "<|endoftext|>",
+  "pad_token": "<|padding|>"
+}

training_meta.json

@@ -0,0 +1,4 @@
+{
+  "words_processed": 92469402,
+  "checkpoint_name": "chck_92M"
+}