Instructions to use ryoji-info/Gemma-4-12B-PsiLM with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use ryoji-info/Gemma-4-12B-PsiLM with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir Gemma-4-12B-PsiLM ryoji-info/Gemma-4-12B-PsiLM
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
2D bridges + config
Browse files
bridges/gemma-4-12b-4bit-mlx-2d-dpot/bridges.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dadf52b9227964ec7db04a1fc6a9c076b6b54faa674b7862f99dccb22889a20b
|
| 3 |
+
size 55302128
|
bridges/gemma-4-12b-4bit-mlx-2d-dpot/config.json
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"backbone": "mlx-community/gemma-4-12B-it-4bit",
|
| 3 |
+
"hf_tokenizer": "mlx-community/gemma-4-12B-it-4bit",
|
| 4 |
+
"task": "2d",
|
| 5 |
+
"loader": "psilm.mlx.gemma_loader.load_backbone_any (gemma4_unified -> text tower)",
|
| 6 |
+
"physics": {
|
| 7 |
+
"file": "physics/dpot_tiny_fisher2d_finetuned.safetensors",
|
| 8 |
+
"dpot_base": "physics/model_Ti.pth",
|
| 9 |
+
"source": "results/stage2d/dpot_ft.pt (DPOT-Tiny fine-tuned on 2D Fisher-KPP)"
|
| 10 |
+
},
|
| 11 |
+
"bridges_class": "psilm.mlx.bridges2d.PsiBridges2DMLX",
|
| 12 |
+
"construct": {
|
| 13 |
+
"d_model": 3840,
|
| 14 |
+
"channel": "value",
|
| 15 |
+
"inj_cap": 0.2,
|
| 16 |
+
"gate_bias": 0.0,
|
| 17 |
+
"readout_norm": "dim"
|
| 18 |
+
},
|
| 19 |
+
"coupling": {
|
| 20 |
+
"l_fwd": 20,
|
| 21 |
+
"l_rev": 30,
|
| 22 |
+
"n_layers": 48
|
| 23 |
+
},
|
| 24 |
+
"pointer": "deterministic span pooling with calibrated per-dimension standardization (fwd.dim_mu/dim_sigma included; fwd.x0_query/x0_key omitted: unused)",
|
| 25 |
+
"training": {
|
| 26 |
+
"steps_total": 7500,
|
| 27 |
+
"phase_A_readout_only": 2000,
|
| 28 |
+
"coupled_steps": 4000,
|
| 29 |
+
"no_harm_steps": 1500,
|
| 30 |
+
"no_harm_resumed_from": 6000,
|
| 31 |
+
"batch": 4,
|
| 32 |
+
"lr": "3e-4 (warm-up, coupled), 1e-4 (no-harm phase)",
|
| 33 |
+
"no_harm_arm": "non-physics prompts (data/noharm_gemma_all.json) paired with the backbone's own greedy continuation; gate-only updates + mean-gate penalty"
|
| 34 |
+
},
|
| 35 |
+
"held_out_n48_per_chunk": {
|
| 36 |
+
"coupled": [
|
| 37 |
+
0.5,
|
| 38 |
+
0.646,
|
| 39 |
+
0.771,
|
| 40 |
+
0.771,
|
| 41 |
+
0.854,
|
| 42 |
+
0.792,
|
| 43 |
+
0.792,
|
| 44 |
+
0.938
|
| 45 |
+
],
|
| 46 |
+
"no_harm_phase": [
|
| 47 |
+
0.979,
|
| 48 |
+
0.958,
|
| 49 |
+
1.0
|
| 50 |
+
]
|
| 51 |
+
},
|
| 52 |
+
"license": "apache-2.0"
|
| 53 |
+
}
|