docs: update default coach-summary model to build-small-hackathon/pozify-coach-summary1 and clarify fallback behavior in documentation

README.md

@@ -46,14 +46,13 @@ The current codebase supports:
 - verifier and conservative fallback summaries
 - Modal training pipelines for both the exercise router and coach-summary model
 
-The current safest cloud runtime for coach summary is:
+The current default coach-summary model is:
 
-- `Qwen/Qwen3-14B`
+- `build-small-hackathon/pozify-coach-summary1`
 
-The current fine-tuned coach-summary Hugging Face repo can be trained, merged, and published, but
-it is not yet cleanly usable through the current Hugging Face serverless `chat_completion` path.
-If you want to use that fine-tuned model today, the simplest path is local merged-model inference
-through `POZIFY_COACH_SUMMARY_LOCAL_MODEL_DIR`.
+Pozify first tries Hugging Face `chat_completion`, then falls back to Hugging Face
+`text_generation` for non-chat model repos. If hosted inference still rejects the repo or returns an
+invalid schema, the app keeps the conservative fallback summary.
 
 ## Product Flow
 
@@ -82,25 +81,25 @@ Supported router labels:
 
 ## Model Stack
 
-| Component | Model or method | Trained here? | Runtime |
-| --- | --- | --- | --- |
-| Pose extraction | MediaPipe Pose Landmarker Lite | No | CPU / MediaPipe delegate |
-| Exercise router | PyTorch BiLSTM over 30-frame pose windows | Yes | Torch |
-| Router baseline | scikit-learn `HistGradientBoostingClassifier` | Yes | CPU fallback/reference |
-| Rep counting | Exercise-specific state machines | No ML | CPU |
-| Issue markers | Transparent rules over per-rep metrics | No ML | CPU |
-| Coach summary | `Qwen/Qwen3-14B` by default | Base model only | HF Inference, local Transformers, or llama.cpp |
-| Coach-summary fine-tune | LoRA / merged checkpoint pipeline on Modal | Yes | Local merged-model path recommended |
-| Verifier | Deterministic safety and grounding checks | No ML | CPU |
+| Component               | Model or method                                          | Trained here?           | Runtime                                        |
+| ----------------------- | -------------------------------------------------------- | ----------------------- | ---------------------------------------------- |
+| Pose extraction         | MediaPipe Pose Landmarker Lite                           | No                      | CPU / MediaPipe delegate                       |
+| Exercise router         | PyTorch BiLSTM over 30-frame pose windows                | Yes                     | Torch                                          |
+| Router baseline         | scikit-learn `HistGradientBoostingClassifier`            | Yes                     | CPU fallback/reference                         |
+| Rep counting            | Exercise-specific state machines                         | No ML                   | CPU                                            |
+| Issue markers           | Transparent rules over per-rep metrics                   | No ML                   | CPU                                            |
+| Coach summary           | `build-small-hackathon/pozify-coach-summary1` by default | Fine-tuned merged model | HF Inference, local Transformers, or llama.cpp |
+| Coach-summary fine-tune | LoRA / merged checkpoint pipeline on Modal               | Yes                     | Local merged-model path recommended            |
+| Verifier                | Deterministic safety and grounding checks                | No ML                   | CPU                                            |
 
 The trained router is intentionally tiny:
 
-| Artifact | Count |
-| --- | ---: |
-| BiLSTM router trainable params | 182,796 |
-| Router input features per frame | 237 |
-| Window length | 30 frames |
-| Output classes | 4 |
+| Artifact                        |     Count |
+| ------------------------------- | --------: |
+| BiLSTM router trainable params  |   182,796 |
+| Router input features per frame |       237 |
+| Window length                   | 30 frames |
+| Output classes                  |         4 |
 
 ## Run The App Locally
 
@@ -146,17 +145,18 @@ uv run python app.py
 
 ## Coach Summary Runtime Options
 
-### 1. Simplest cloud path
+### 1. Fine-tuned coach model
 
-Use the default supported cloud model:
+The app defaults to the fine-tuned coach-summary model:
 
 ```bash
-export POZIFY_COACH_SUMMARY_MODEL=Qwen/Qwen3-14B
+export POZIFY_COACH_SUMMARY_MODEL=build-small-hackathon/pozify-coach-summary1
 uv run python app.py
 ```
 
-This is the easiest path if you want the UI to avoid immediate fallback caused by unsupported custom
-HF serverless routing.
+Pozify tries `chat_completion` first and falls back to `text_generation` when Hugging Face reports
+that the repo is not a chat model. The deterministic fallback summary remains enabled if hosted
+inference is unavailable or the model output fails validation.
 
 ### 2. Use the fine-tuned merged model locally
 
@@ -172,7 +172,16 @@ uv run python app.py
 This is the simplest way to use `build-small-hackathon/pozify-coach-summary1` today without adding a
 dedicated inference endpoint.
 
-### 3. llama.cpp
+### 3. Base cloud model override
+
+If you need the previous base-model runtime:
+
+```bash
+export POZIFY_COACH_SUMMARY_MODEL=Qwen/Qwen3-14B
+uv run python app.py
+```
+
+### 4. llama.cpp
 
 Pozify can send the coach-summary prompt to a local `llama-server` that exposes the
 OpenAI-compatible `/v1/chat/completions` endpoint.
@@ -200,14 +209,14 @@ uv run python app.py
 
 ### Useful environment variables
 
-| Variable | Purpose |
-| --- | --- |
-| `POZIFY_ROUTER_DEVICE` | Override router device, for example `cpu` or `cuda`. |
-| `POZIFY_SPACES_GPU_DURATION` | `spaces.GPU` duration in seconds, default `120`. |
-| `POZIFY_COACH_SUMMARY_PROVIDER` | `hf_inference`, `local_transformers`, or `llama_cpp`. |
-| `POZIFY_COACH_SUMMARY_MODEL` | Coach model id or llama.cpp model alias. |
+| Variable                               | Purpose                                                  |
+| -------------------------------------- | -------------------------------------------------------- |
+| `POZIFY_ROUTER_DEVICE`                 | Override router device, for example `cpu` or `cuda`.     |
+| `POZIFY_SPACES_GPU_DURATION`           | `spaces.GPU` duration in seconds, default `120`.         |
+| `POZIFY_COACH_SUMMARY_PROVIDER`        | `hf_inference`, `local_transformers`, or `llama_cpp`.    |
+| `POZIFY_COACH_SUMMARY_MODEL`           | Coach model id or llama.cpp model alias.                 |
 | `POZIFY_COACH_SUMMARY_LOCAL_MODEL_DIR` | Prefer a local merged/model directory for coach summary. |
-| `POZIFY_COACH_SUMMARY_BYPASS_VERIFIER` | Keep model output even when verifier fails. |
+| `POZIFY_COACH_SUMMARY_BYPASS_VERIFIER` | Keep model output even when verifier fails.              |
 
 ## Exercise Router Training
 
@@ -262,10 +271,10 @@ uv run modal run scripts/coach_summary_modal.py --stage publish-merged --repo-id
 
 Important runtime note:
 
-- publishing the merged repo does not automatically make it usable through HF serverless
-  `chat_completion`
-- if you want immediate working inference, use `Qwen/Qwen3-14B`
-- if you want to use the fine-tuned model, use `POZIFY_COACH_SUMMARY_LOCAL_MODEL_DIR`
+- the default coach model is `build-small-hackathon/pozify-coach-summary1`
+- Hugging Face hosted inference may still reject a repo or produce invalid JSON, so the
+  conservative fallback summary stays enabled
+- for the most predictable fine-tuned inference path, use `POZIFY_COACH_SUMMARY_LOCAL_MODEL_DIR`
 
 ## Generated Artifacts
 

docs/01-docs-index.md

@@ -20,7 +20,7 @@ Read in this order if you want to understand and run the current project:
 
 ## Notes
 
-- Current coach-summary cloud runtime default: `Qwen/Qwen3-14B`
-- Current fine-tuned coach-summary Hugging Face repo is best used through local merged-model
-  inference, not HF serverless `chat_completion`
+- Current coach-summary runtime default: `build-small-hackathon/pozify-coach-summary1`
+- The Hugging Face provider tries `chat_completion` first, then `text_generation` for non-chat
+  model repos; local merged-model inference remains the most predictable fine-tuned path
 - Current app entrypoint: `uv run python app.py`

docs/10-overview-build-small-hackathon-report.md

@@ -3,7 +3,7 @@
 Status note:
 
 - This report is kept as the hackathon narrative document.
-- Current runtime defaults have moved to `Qwen/Qwen3-14B`.
+- Current runtime defaults use `build-small-hackathon/pozify-coach-summary1`.
 - Current coach-summary training now includes a LoRA/merge/publish pipeline on Modal.
 - For the current operational commands, prefer [../README.md](../README.md) and
   [30-coach-modal-training.md](30-coach-modal-training.md).
@@ -57,7 +57,7 @@ pretending every video is one of the supported movements.
 | Pose extractor | MediaPipe Pose Landmarker Lite | Fast, practical feature extractor for a Gradio Space. |
 | Exercise router | Custom PyTorch BiLSTM | Tiny trainable temporal model over pose windows. |
 | Baseline router | scikit-learn HistGradientBoostingClassifier | Strong baseline over engineered vectors and fallback artifact. |
-| Coach summary | Qwen/Qwen3-14B | Current default cloud runtime for structured JSON explanation. |
+| Coach summary | build-small-hackathon/pozify-coach-summary1 | Current default fine-tuned runtime for structured JSON explanation. |
 | llama.cpp path | Qwen3-14B Instruct GGUF via `llama-server` | Local-first/off-grid coach summary path with GPU offload. |
 
 The original hackathon build trained the exercise router first and used Qwen as a grounded
@@ -212,7 +212,7 @@ logic stay outside the GPU worker. Useful settings:
 
 ```bash
 POZIFY_COACH_SUMMARY_PROVIDER=local_transformers
-POZIFY_COACH_SUMMARY_MODEL=Qwen/Qwen3-14B
+POZIFY_COACH_SUMMARY_MODEL=build-small-hackathon/pozify-coach-summary1
 POZIFY_SPACES_GPU_DURATION=300
 ```
 

docs/30-coach-modal-training.md

@@ -34,8 +34,9 @@ Important limitation:
 
 - Publishing a merged model repo to Hugging Face does **not** currently guarantee that the repo can
   be used through Hugging Face serverless `chat_completion`.
-- In the current codebase, the safest cloud runtime remains `Qwen/Qwen3-14B`.
-- The simplest way to use the fine-tuned merged model today is local inference through
+- The current codebase defaults to `build-small-hackathon/pozify-coach-summary1` and falls back
+  from `chat_completion` to `text_generation` for non-chat model repos.
+- The most predictable fine-tuned path remains local inference through
   `POZIFY_COACH_SUMMARY_LOCAL_MODEL_DIR`.
 
 ## Requirements
@@ -142,10 +143,10 @@ The Modal model volume stores:
 
 ## Current Runtime Usage
 
-### Safest cloud runtime now
+### Default fine-tuned runtime
 
 ```bash
-export POZIFY_COACH_SUMMARY_MODEL=Qwen/Qwen3-14B
+export POZIFY_COACH_SUMMARY_MODEL=build-small-hackathon/pozify-coach-summary1
 uv run python app.py
 ```
 
@@ -160,15 +161,16 @@ export POZIFY_COACH_SUMMARY_ADAPTER_ID=build-small-hackathon/pozify-coach-summar
 uv run python app.py
 ```
 
-### Not recommended right now
+### Previous base-model runtime
 
-Pointing app runtime directly at the merged Hugging Face repo through:
+If you need to force the previous base model:
 
 ```bash
-export POZIFY_COACH_SUMMARY_MODEL=build-small-hackathon/pozify-coach-summary1
+export POZIFY_COACH_SUMMARY_MODEL=Qwen/Qwen3-14B
 ```
 
-This may still fall back because HF serverless currently rejects that repo as “not a chat model”.
+The fine-tuned model may still fall back if hosted inference is unavailable or returns invalid JSON,
+so the deterministic fallback summary remains part of the runtime path.
 
 ## Evaluation Meaning
 

src/pozify/slm/providers.py

@@ -24,7 +24,7 @@ LLAMA_CPP_BASE_URL_ENV = "POZIFY_LLAMA_CPP_BASE_URL"
 LLAMA_CPP_TIMEOUT_ENV = "POZIFY_LLAMA_CPP_TIMEOUT"
 
 DEFAULT_PROVIDER = "hf_inference"
-DEFAULT_MODEL = "Qwen/Qwen3-14B"
+DEFAULT_MODEL = "build-small-hackathon/pozify-coach-summary1"
 DEFAULT_LLAMA_CPP_BASE_URL = "http://127.0.0.1:8080"
 LOCAL_TRANSFORMERS_PROVIDER = "local_transformers"
 LOCAL_TRANSFORMERS_ALIASES = {LOCAL_TRANSFORMERS_PROVIDER, "local", "transformers"}
@@ -97,52 +97,83 @@ class HFInferenceCoachSummaryModel:
         self._client = InferenceClient(api_key=self.token)
         return self._client
 
+    def _generate_with_chat_completion(
+        self,
+        client: Any,
+        prompt: str,
+    ) -> CoachSummaryGeneration:
+        response = client.chat_completion(
+            model=self.model,
+            messages=[
+                {
+                    "role": "system",
+                    "content": "Return JSON only.",
+                },
+                {
+                    "role": "user",
+                    "content": prompt,
+                },
+            ],
+            temperature=self.temperature,
+            max_tokens=self.max_tokens,
+        )
+        choices = getattr(response, "choices", None) or []
+        if not choices:
+            raise RuntimeError("Hugging Face inference returned no choices")
+        message = getattr(choices[0], "message", None)
+        text = getattr(message, "content", None)
+        if not isinstance(text, str) or not text.strip():
+            raise RuntimeError("Hugging Face inference returned an empty message")
+        return CoachSummaryGeneration(
+            text=text,
+            provider="hf_inference",
+            model=self.model,
+        )
+
+    def _generate_with_text_generation(
+        self,
+        client: Any,
+        prompt: str,
+    ) -> CoachSummaryGeneration:
+        text = client.text_generation(
+            prompt,
+            model=self.model,
+            max_new_tokens=self.max_tokens,
+            temperature=self.temperature,
+            return_full_text=False,
+        )
+        if not isinstance(text, str) or not text.strip():
+            raise RuntimeError("Hugging Face text generation returned empty output")
+        return CoachSummaryGeneration(
+            text=text,
+            provider="hf_inference",
+            model=self.model,
+        )
+
     def generate_summary(self, prompt: str) -> CoachSummaryGeneration:
         client = self._client_instance()
+        chat_error: Exception | None = None
         if hasattr(client, "chat_completion"):
-            response = client.chat_completion(
-                model=self.model,
-                messages=[
-                    {
-                        "role": "system",
-                        "content": "Return JSON only.",
-                    },
-                    {
-                        "role": "user",
-                        "content": prompt,
-                    },
-                ],
-                temperature=self.temperature,
-                max_tokens=self.max_tokens,
-            )
-            choices = getattr(response, "choices", None) or []
-            if not choices:
-                raise RuntimeError("Hugging Face inference returned no choices")
-            message = getattr(choices[0], "message", None)
-            text = getattr(message, "content", None)
-            if not isinstance(text, str) or not text.strip():
-                raise RuntimeError("Hugging Face inference returned an empty message")
-            return CoachSummaryGeneration(
-                text=text,
-                provider="hf_inference",
-                model=self.model,
-            )
+            try:
+                return self._generate_with_chat_completion(client, prompt)
+            except Exception as exc:
+                chat_error = exc
 
         if hasattr(client, "text_generation"):
-            text = client.text_generation(
-                prompt,
-                model=self.model,
-                max_new_tokens=self.max_tokens,
-                temperature=self.temperature,
-                return_full_text=False,
-            )
-            if not isinstance(text, str) or not text.strip():
-                raise RuntimeError("Hugging Face text generation returned empty output")
-            return CoachSummaryGeneration(
-                text=text,
-                provider="hf_inference",
-                model=self.model,
-            )
+            try:
+                return self._generate_with_text_generation(client, prompt)
+            except Exception as exc:
+                if chat_error is None:
+                    raise
+                raise RuntimeError(
+                    "Hugging Face chat completion and text generation both failed: "
+                    f"chat={chat_error}; text_generation={exc}"
+                ) from exc
+
+        if chat_error is not None:
+            raise RuntimeError(
+                f"Hugging Face chat completion failed and text generation is unavailable: {chat_error}"
+            ) from chat_error
 
         raise RuntimeError("No supported Hugging Face inference method is available")
 

tests/test_slm_providers.py

@@ -67,7 +67,43 @@ class SlmProviderTests(unittest.TestCase):
             model = get_coach_summary_model()
 
         self.assertIsInstance(model, HFInferenceCoachSummaryModel)
-        self.assertEqual(model.model, "Qwen/Qwen3-14B")
+        self.assertEqual(model.model, "build-small-hackathon/pozify-coach-summary1")
+
+    def test_hf_inference_falls_back_to_text_generation_for_non_chat_model(self) -> None:
+        class _TextGenerationClient:
+            def __init__(self) -> None:
+                self.text_generation_kwargs = None
+
+            def chat_completion(self, **_kwargs):
+                raise RuntimeError("not a chat model")
+
+            def text_generation(self, prompt: str, **kwargs):
+                self.text_generation_kwargs = {"prompt": prompt, **kwargs}
+                return '{"summary":"ok"}'
+
+        client = _TextGenerationClient()
+        model = HFInferenceCoachSummaryModel(
+            model="build-small-hackathon/pozify-coach-summary1",
+            max_tokens=123,
+            temperature=0.2,
+        )
+        model._client = client
+
+        generation = model.generate_summary("coach prompt")
+
+        self.assertEqual(generation.provider, "hf_inference")
+        self.assertEqual(generation.model, "build-small-hackathon/pozify-coach-summary1")
+        self.assertEqual(generation.text, '{"summary":"ok"}')
+        self.assertEqual(
+            client.text_generation_kwargs,
+            {
+                "prompt": "coach prompt",
+                "model": "build-small-hackathon/pozify-coach-summary1",
+                "max_new_tokens": 123,
+                "temperature": 0.2,
+                "return_full_text": False,
+            },
+        )
 
 
 if __name__ == "__main__":