Upload folder using huggingface_hub

README.md

@@ -0,0 +1,46 @@
+---
+license: apache-2.0
+base_model: mlx-community/Qwen3-1.7B-4bit
+base_model_relation: adapter
+library_name: mlx
+pipeline_tag: text-generation
+datasets:
+  - caiotheodoro/plumb
+tags:
+  - mlx
+  - lora
+  - construction
+  - pay-application
+  - curriculum
+---
+
+# plumb-handseeded
+
+1.7B MLX LoRA trained on the **hand-seeded** Path A curriculum (223 tasks, distribution-matched). Same oracle and seed-777 eval as the Ornith and blend adapters. Only the training mix changes.
+
+| | |
+|---|---|
+| Base | [`mlx-community/Qwen3-1.7B-4bit`](https://huggingface.co/mlx-community/Qwen3-1.7B-4bit) |
+| Train | 190 / 8 epochs / val loss 0.486 / ~3.3GB on an M5 |
+| Gold | [`caiotheodoro/plumb`](https://huggingface.co/datasets/caiotheodoro/plumb) `train_handseeded` |
+| Code | [caiotheodoro/plumb](https://github.com/caiotheodoro/plumb) |
+
+## Result (n=1000, seed-777)
+
+| | sw-recall | precision | exact | parse |
+|---|---|---|---|---|
+| **this adapter** | **0.318 [0.290, 0.347]** | 0.308 [0.279, 0.337] | 0.178 | 1.000 |
+| Ornith-only | 0.241 [0.214, 0.268] | 0.111 [0.098, 0.124] | 0.084 | 0.997 |
+| blend | 0.334 [0.306, 0.363] | 0.374 [0.342, 0.406] | 0.228 | 1.000 |
+
+Ornith-only sits below this row with no overlap on recall. Blend overlaps this row on recall; precision is the interval that survives.
+
+Text policy, not 27B VL. Synthetic G702. N=18 mix/error runs are a null and are not these weights.
+
+```python
+from huggingface_hub import snapshot_download
+from mlx_lm.lora import load
+
+path = snapshot_download("caiotheodoro/plumb-handseeded")
+model, tokenizer = load("mlx-community/Qwen3-1.7B-4bit", adapter_path=path)
+```

adapter_config.json

@@ -0,0 +1,41 @@
+{
+    "adapter_path": "adapters/text-handseeded",
+    "batch_size": 1,
+    "clear_cache_threshold": 1000000000,
+    "data": "data/text-handseeded",
+    "fine_tune_type": "lora",
+    "grad_accumulation_steps": 1,
+    "grad_checkpoint": true,
+    "iters": 1520,
+    "learning_rate": 1e-05,
+    "lora_parameters": {
+        "rank": 16,
+        "dropout": 0.05,
+        "scale": 2.0
+    },
+    "lr_schedule": null,
+    "mask_prompt": false,
+    "max_seq_length": 2048,
+    "model": "mlx-community/Qwen3-1.7B-4bit",
+    "num_layers": 16,
+    "optimizer": "adamw",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+        "muon": {},
+        "sgd": {},
+        "adafactor": {}
+    },
+    "project_name": null,
+    "quantization": null,
+    "quantization_config": null,
+    "report_to": null,
+    "resume_adapter_file": null,
+    "save_every": 50,
+    "seed": 7,
+    "steps_per_eval": 50,
+    "steps_per_report": 10,
+    "test": false,
+    "train": true,
+    "val_batches": -1
+}

adapters.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f0c765947efb88966f9a9b73033452a5825f7f44f7254c493f86121b4e9357a
+size 39870300

train_summary.json

@@ -0,0 +1,10 @@
+{
+  "base": "mlx-community/Qwen3-1.7B-4bit",
+  "adapter_path": "adapters/text-handseeded",
+  "iters": 1520,
+  "epochs": 8,
+  "n_train": 190,
+  "final_val_loss": 0.4860517680644989,
+  "wall_seconds": 5209.0,
+  "peak_memory_gb": 3.34
+}