Stage validated dataset and base model during free build

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* 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
+data/dataset/train.jsonl filter=lfs diff=lfs merge=lfs -text
+prepared_dataset/train.jsonl filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -19,6 +19,10 @@ COPY requirements.txt .
 RUN python -m pip install --no-cache-dir --upgrade pip && \
     python -m pip install --no-cache-dir -r requirements.txt
 
+# Cache the public base model while the Space image is building. Hugging Face
+# does not bill requested GPU hardware during the build stage.
+RUN python -c "from huggingface_hub import snapshot_download; snapshot_download('Qwen/Qwen3-1.7B')"
+
 # Pinned official llama.cpp revision for reproducible GGUF conversion.
 RUN git clone https://github.com/ggml-org/llama.cpp.git /opt/llama.cpp && \
     cd /opt/llama.cpp && \

data/dataset/manifest.json

@@ -0,0 +1,37 @@
+{
+  "actual_examples": 7925,
+  "categories": {
+    "activity_context": 96,
+    "adversarial_input": 500,
+    "altitude_telemetry": 23,
+    "communications_loss": 400,
+    "missing_stale_data": 800,
+    "mountain_knowledge": 50,
+    "multi_turn_checkin": 600,
+    "noisy_speech": 600,
+    "telemetry_safety": 4748,
+    "wearable_reliability": 108
+  },
+  "notes": [
+    "Harespod values are treated as normalized research signals, not clinical units.",
+    "MedSafetyBench is excluded from SFT and reserved for evaluation.",
+    "Synthetic records are policy-grounded and explicitly marked in metadata.",
+    "Splits are grouped by source subject or scenario family to reduce leakage."
+  ],
+  "ppg_dalia_available": false,
+  "seed": 42,
+  "sources": {
+    "cdc-yellow-book-altitude-2026": 45,
+    "harespod": 23,
+    "mhealth": 108,
+    "pamap2": 96,
+    "synthetic-policy-grounded": 7648,
+    "wms-wc24-altitude-syllabus": 5
+  },
+  "splits": {
+    "test": 714,
+    "train": 6489,
+    "validation": 722
+  },
+  "target_examples": 8000
+}

data/dataset/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:452b92ef1a1f2c6bd65027d1688a3ec999410046284278342a7f31adbe735a7e
+size 10895160

data/dataset/validation_report.json

@@ -0,0 +1,30 @@
+{
+  "actions": {
+    "ACTIVATE_EMERGENCY": 751,
+    "CONTINUE_MONITORING": 666,
+    "HOLD_ASCENT": 2318,
+    "RECHECK_SENSOR": 3767,
+    "REQUEST_HUMAN_REVIEW": 423
+  },
+  "categories": {
+    "activity_context": 96,
+    "adversarial_input": 500,
+    "altitude_telemetry": 23,
+    "communications_loss": 400,
+    "missing_stale_data": 800,
+    "mountain_knowledge": 50,
+    "multi_turn_checkin": 600,
+    "noisy_speech": 600,
+    "telemetry_safety": 4748,
+    "wearable_reliability": 108
+  },
+  "counts": {
+    "test": 714,
+    "train": 6489,
+    "validation": 722
+  },
+  "errors": [],
+  "ok": true,
+  "unique_groups": 2443,
+  "warnings": []
+}

data/logs/events.jsonl

@@ -0,0 +1,4 @@
+{"actual_examples": 8000, "categories": {"adversarial_input": 500, "communications_loss": 400, "missing_stale_data": 800, "mountain_knowledge": 50, "multi_turn_checkin": 600, "noisy_speech": 600, "telemetry_safety": 5050}, "event": "dataset_built", "notes": ["Harespod values are treated as normalized research signals, not clinical units.", "MedSafetyBench is excluded from SFT and reserved for evaluation.", "Synthetic records are policy-grounded and explicitly marked in metadata.", "Splits are grouped by source subject or scenario family to reduce leakage."], "ppg_dalia_available": false, "seed": 42, "sources": {"cdc-yellow-book-altitude-2026": 45, "synthetic-policy-grounded": 7950, "wms-wc24-altitude-syllabus": 5}, "splits": {"test": 708, "train": 6542, "validation": 750}, "target_examples": 8000, "time": "2026-08-29T23:49:51.358270+00:00"}
+{"actions": {"ACTIVATE_EMERGENCY": 768, "CONTINUE_MONITORING": 529, "HOLD_ASCENT": 2419, "RECHECK_SENSOR": 3884, "REQUEST_HUMAN_REVIEW": 400}, "categories": {"adversarial_input": 500, "communications_loss": 400, "missing_stale_data": 800, "mountain_knowledge": 50, "multi_turn_checkin": 600, "noisy_speech": 600, "telemetry_safety": 5050}, "counts": {"test": 708, "train": 6542, "validation": 750}, "errors": [], "event": "dataset_validated", "ok": true, "time": "2026-08-29T23:49:52.217503+00:00", "unique_groups": 2490, "warnings": []}
+{"event": "prepared_dataset_staged", "files": ["train.jsonl", "validation.jsonl", "test.jsonl", "manifest.json", "validation_report.json"], "time": "2026-08-30T01:08:33.736347+00:00"}
+{"actions": {"ACTIVATE_EMERGENCY": 751, "CONTINUE_MONITORING": 666, "HOLD_ASCENT": 2318, "RECHECK_SENSOR": 3767, "REQUEST_HUMAN_REVIEW": 423}, "categories": {"activity_context": 96, "adversarial_input": 500, "altitude_telemetry": 23, "communications_loss": 400, "missing_stale_data": 800, "mountain_knowledge": 50, "multi_turn_checkin": 600, "noisy_speech": 600, "telemetry_safety": 4748, "wearable_reliability": 108}, "counts": {"test": 714, "train": 6489, "validation": 722}, "errors": [], "event": "dataset_validated", "ok": true, "time": "2026-08-30T01:08:34.572972+00:00", "unique_groups": 2443, "warnings": []}

data/status/status.json

@@ -0,0 +1,67 @@
+{
+  "dataset": {
+    "actual_examples": 8000,
+    "categories": {
+      "adversarial_input": 500,
+      "communications_loss": 400,
+      "missing_stale_data": 800,
+      "mountain_knowledge": 50,
+      "multi_turn_checkin": 600,
+      "noisy_speech": 600,
+      "telemetry_safety": 5050
+    },
+    "notes": [
+      "Harespod values are treated as normalized research signals, not clinical units.",
+      "MedSafetyBench is excluded from SFT and reserved for evaluation.",
+      "Synthetic records are policy-grounded and explicitly marked in metadata.",
+      "Splits are grouped by source subject or scenario family to reduce leakage."
+    ],
+    "ppg_dalia_available": false,
+    "seed": 42,
+    "sources": {
+      "cdc-yellow-book-altitude-2026": 45,
+      "synthetic-policy-grounded": 7950,
+      "wms-wc24-altitude-syllabus": 5
+    },
+    "splits": {
+      "test": 708,
+      "train": 6542,
+      "validation": 750
+    },
+    "target_examples": 8000
+  },
+  "message": "Dataset validation passed",
+  "phase": "dataset_valid",
+  "pid": 39260,
+  "updated_at": "2026-08-30T01:08:34.574191+00:00",
+  "validation": {
+    "actions": {
+      "ACTIVATE_EMERGENCY": 751,
+      "CONTINUE_MONITORING": 666,
+      "HOLD_ASCENT": 2318,
+      "RECHECK_SENSOR": 3767,
+      "REQUEST_HUMAN_REVIEW": 423
+    },
+    "categories": {
+      "activity_context": 96,
+      "adversarial_input": 500,
+      "altitude_telemetry": 23,
+      "communications_loss": 400,
+      "missing_stale_data": 800,
+      "mountain_knowledge": 50,
+      "multi_turn_checkin": 600,
+      "noisy_speech": 600,
+      "telemetry_safety": 4748,
+      "wearable_reliability": 108
+    },
+    "counts": {
+      "test": 714,
+      "train": 6489,
+      "validation": 722
+    },
+    "errors": [],
+    "ok": true,
+    "unique_groups": 2443,
+    "warnings": []
+  }
+}

prepared_dataset/manifest.json

@@ -0,0 +1,37 @@
+{
+  "actual_examples": 7925,
+  "categories": {
+    "activity_context": 96,
+    "adversarial_input": 500,
+    "altitude_telemetry": 23,
+    "communications_loss": 400,
+    "missing_stale_data": 800,
+    "mountain_knowledge": 50,
+    "multi_turn_checkin": 600,
+    "noisy_speech": 600,
+    "telemetry_safety": 4748,
+    "wearable_reliability": 108
+  },
+  "notes": [
+    "Harespod values are treated as normalized research signals, not clinical units.",
+    "MedSafetyBench is excluded from SFT and reserved for evaluation.",
+    "Synthetic records are policy-grounded and explicitly marked in metadata.",
+    "Splits are grouped by source subject or scenario family to reduce leakage."
+  ],
+  "ppg_dalia_available": false,
+  "seed": 42,
+  "sources": {
+    "cdc-yellow-book-altitude-2026": 45,
+    "harespod": 23,
+    "mhealth": 108,
+    "pamap2": 96,
+    "synthetic-policy-grounded": 7648,
+    "wms-wc24-altitude-syllabus": 5
+  },
+  "splits": {
+    "test": 714,
+    "train": 6489,
+    "validation": 722
+  },
+  "target_examples": 8000
+}

prepared_dataset/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:452b92ef1a1f2c6bd65027d1688a3ec999410046284278342a7f31adbe735a7e
+size 10895160

prepared_dataset/validation_report.json

@@ -0,0 +1,30 @@
+{
+  "actions": {
+    "ACTIVATE_EMERGENCY": 751,
+    "CONTINUE_MONITORING": 666,
+    "HOLD_ASCENT": 2318,
+    "RECHECK_SENSOR": 3767,
+    "REQUEST_HUMAN_REVIEW": 423
+  },
+  "categories": {
+    "activity_context": 96,
+    "adversarial_input": 500,
+    "altitude_telemetry": 23,
+    "communications_loss": 400,
+    "missing_stale_data": 800,
+    "mountain_knowledge": 50,
+    "multi_turn_checkin": 600,
+    "noisy_speech": 600,
+    "telemetry_safety": 4748,
+    "wearable_reliability": 108
+  },
+  "counts": {
+    "test": 714,
+    "train": 6489,
+    "validation": 722
+  },
+  "errors": [],
+  "ok": true,
+  "unique_groups": 2443,
+  "warnings": []
+}

trainer/run_pipeline.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import os
+import shutil
 import subprocess
 import sys
 import traceback
@@ -8,7 +9,17 @@ from pathlib import Path
 
 from huggingface_hub import HfApi
 
-from .common import ARTIFACT_ROOT, STATUS_ROOT, ensure_dirs, fail, log_event, update_status, wait_for_gpu
+from .common import (
+    APP_ROOT,
+    ARTIFACT_ROOT,
+    DATASET_ROOT,
+    STATUS_ROOT,
+    ensure_dirs,
+    fail,
+    log_event,
+    update_status,
+    wait_for_gpu,
+)
 
 
 STAGES = [
@@ -21,6 +32,15 @@ STAGES = [
     ("merge_export", "trainer.merge_and_export"),
 ]
 
+PREPARED_DATASET_ROOT = APP_ROOT / "prepared_dataset"
+PREPARED_DATASET_FILES = (
+    "train.jsonl",
+    "validation.jsonl",
+    "test.jsonl",
+    "manifest.json",
+    "validation_report.json",
+)
+
 
 def run_module(name: str, extra_env: dict[str, str] | None = None) -> None:
     env = os.environ.copy()
@@ -28,6 +48,20 @@ def run_module(name: str, extra_env: dict[str, str] | None = None) -> None:
     subprocess.run([sys.executable, "-m", name], check=True, env=env)
 
 
+def stage_prepared_dataset() -> bool:
+    if not PREPARED_DATASET_ROOT.is_dir():
+        return False
+    missing = [name for name in PREPARED_DATASET_FILES if not (PREPARED_DATASET_ROOT / name).is_file()]
+    if missing:
+        raise RuntimeError(f"Bundled prepared dataset is incomplete: {', '.join(missing)}")
+    update_status("stage_dataset", "Staging the bundled, locally validated SFT dataset")
+    DATASET_ROOT.mkdir(parents=True, exist_ok=True)
+    for name in PREPARED_DATASET_FILES:
+        shutil.copy2(PREPARED_DATASET_ROOT / name, DATASET_ROOT / name)
+    log_event("prepared_dataset_staged", files=list(PREPARED_DATASET_FILES))
+    return True
+
+
 def upload_artifacts() -> str:
     token = os.environ.get("TRAINING_HF_TOKEN")
     repo_id = os.environ.get("OUTPUT_MODEL_REPO", "iteratehack/sam-qwen3-1.7b-mountain")
@@ -78,8 +112,9 @@ def main() -> None:
     lock.write_text(str(os.getpid()), encoding="utf-8")
     try:
         update_status("starting", "Pipeline started")
-        run_module("trainer.download_sources")
-        run_module("trainer.build_dataset")
+        if not stage_prepared_dataset():
+            run_module("trainer.download_sources")
+            run_module("trainer.build_dataset")
         run_module("trainer.validate_dataset")
         wait_for_gpu()