feat(pipelines): add paedawen finetune

pipelines/paedawen/config.yaml

@@ -0,0 +1,38 @@
+# Default fine-tuning config (training params ONLY).
+#
+# Each trainer translates this into native form via config.py
+# Operational args (instance type, AWS config, work_dir, quantize, ...) are NOT here
+# they stay per-trainer constructor inputs because they differ between HF and MLX.
+#
+# Cross-library translation reference:
+#   lora.alpha + lora.rank  ->  mlx scale = alpha / rank   (e.g. 64 / 32 = 2.0)
+#   lora.target_modules     ->  mlx keys  = ["self_attn.<mod>" / "mlp.<mod>", ...]
+#   epochs                  ->  mlx iters = ceil(num_samples / batch_size) * epochs
+#                               (only set mlx.iters for override of derived value)
+
+training:
+  base_model: Qwen/Qwen3-4B-Instruct-2507   # which pre-trained model to start from
+  epochs: 3                         # how many full passes over the training data
+  learning_rate: 2e-4               # how big a step the model takes each update (too high = unstable, too low = slow)
+  batch_size: 1                     # samples processed at once; keep at 1 for long (16k-32k) samples to avoid running out of memory
+  max_seq_length: 16384             # longest sample (in tokens) the model will read - anything longer is cut off
+  lora:
+    rank: 32                        # size of the small trainable add-on; higher = more capacity to learn but more memory
+    alpha: 64                       # how strongly the add-on influences the model (mlx scale = alpha / rank = 2.0)
+    dropout: 0.05                   # randomly ignores 5% of connections while training to reduce overfitting
+    target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]   # which parts of each layer get the trainable add-on
+
+  mlx:                              # optional; consumed ONLY by the MLX translation
+    iters: null                    # total training steps; left null to auto-derive from epochs, set a number to override
+    num_layers: 16                 # how many of the model's top layers to fine-tune - 16 is good balance, -1 means all layers
+    seed: 42                       # fixed random seed so runs are reproducible
+    save_every: 1000                # save a checkpoint every N steps
+    steps_per_report: 10           # print training progress every N steps
+    steps_per_eval: 200            # check performance on held-out data every N steps
+    val_batches: 0                # 0 = no validation (TrainingDataHandler writes train.jsonl only)
+    grad_accumulation_steps: 4    # micro-steps accumulated before each optimizer update
+    grad_checkpoint: true         # trade compute for memory by recomputing activations
+    weight_decay: 0.01            # AdamW weight decay; null uses the mlx default
+    lr_scheduler_type: cosine     # decay the learning rate to 0 over training; null keeps it constant
+    warmup_ratio: 0.05            # fraction of iters spent warming up to learning_rate
+    lr_schedule: null              # raw mlx_lm passthrough for a custom schedule; overrides lr_scheduler_type

pipelines/paedawen/finetune.ipynb

@@ -0,0 +1,96 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "11a389f2",
+   "metadata": {},
+   "source": [
+    "### Paedawen: Model fine-tuning"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4cc82e36",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "from dotenv import load_dotenv\n",
+    "load_dotenv()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "950d260a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from mesa_runtime import MesaRuntime\n",
+    "runtime: MesaRuntime = MesaRuntime(\n",
+    "    os.environ[\"BASE_URL\"], os.environ[\"USERNAME\"], os.environ[\"PASSWORD\"]\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b477ce37",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "runtime.list_training_data()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "7567abd5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "run_id: int = runtime.dispatch(\n",
+    "    \"20260720-223834_paedawen-batch\",\n",
+    "    model_name=\"paedawen\",\n",
+    "    schema=\"paedacute\",\n",
+    "    description=\"Extracting acute triage signs from paediatric triage notes\",\n",
+    "    version=\"0.1.0\",\n",
+    "    config=\"config.yaml\",\n",
+    "    config_from_file=True\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ad3a0857",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "runtime.get_run(run_id=run_id)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "paedawen (3.13.14.final.0)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.13.14"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

pipelines/paedawen/pyproject.toml

@@ -9,8 +9,10 @@ dependencies = [
     "londonaicentre-docsynth",
     "londonaicentre-mesa-datagen",
     "londonaicentre-paedacuteschema>=1.2.0",
+    "mesa-runtime",
 ]
 
 [tool.uv.sources]
 londonaicentre-docsynth = { git = "ssh://git@github.com/londonaicentre/MESA-Docsynth", branch = "spike/mesa-build-refactor" }
 londonaicentre-mesa-datagen = { path = "../../lib/datagen", editable = true }
+mesa-runtime = { git = "https://github.com/londonaicentre/mesa-runtime.git", subdirectory = "lib" }