Simplify setup and link the training tutorial and model weights

.github/workflows/checks.yml

@@ -13,6 +13,6 @@ jobs:
       - name: Install locked dependencies
         run: |
           python -m pip install uv
-          uv sync --locked --extra train
+          uv sync --locked
       - name: Offline unit tests and current run evidence
-        run: uv run --extra train python -m unittest -v
+        run: uv run python -m unittest -v

CONTRIBUTING.md

@@ -1,6 +1,6 @@
 # Contributing
 
-Use Python 3.12+ and `uv sync --locked --extra train`. Run
+Use Python 3.12+ and `uv sync --locked`. Run
 `uv run python -m unittest -v` before submitting changes. Follow the
 [dataset guide](docs/DATASET.md) for the separate data-build and validation steps.
 

README.md

@@ -2,12 +2,14 @@
 
 We fine-tuned **Qwen3.5-4B on Together AI** to make decisions: give it context,
 a question, and 2–24 options, and it returns one answer letter. We're calling
-it **tev1-4B-experimental**. A separate blog post will cover the project, how we
-fine-tuned it, and what we learned.
+it **tev1-4B-experimental**.
 
 This repo contains the data recipe, training example, and saved results so you
 can fine-tune your own decision model.
 
+- [Learn how to train your own classifier for $17](https://www.together.ai/blog/how-to-train-your-own-jev)
+- [Full model weights](https://huggingface.co/togethercomputer/Tev1-4B-experimental)
+
 ```text
 State       Returns are allowed within 30 days. This purchase was 12 days ago.
 Question    Is this return within the allowed window?
@@ -40,7 +42,7 @@ Requires Python 3.12+ and [uv](https://docs.astral.sh/uv/).
 Run these commands from the repository root:
 
 ```bash
-uv sync --locked --extra train
+uv sync --locked
 uv run python fetch_sources.py
 uv run python build_all.py
 ```
@@ -54,9 +56,9 @@ for validation, provenance, and how to adapt the recipe.
 Preview the training settings, then launch when ready:
 
 ```bash
-uv run --extra train python examples/train_together.py
+uv run python examples/train_together.py
 export TOGETHER_API_KEY='your-key'
-uv run --extra train python examples/train_together.py --launch
+uv run python examples/train_together.py --launch
 ```
 
 `--launch` uploads `data/new-v1/instruction/train.jsonl` and `dev.jsonl` and starts

docs/TRAINING.md

@@ -17,8 +17,8 @@ verified export of the historical job's settings.
 | Early stopping / seed | Disabled / 42 |
 | Train on inputs | False; completion-only loss |
 
-Install the locked SDK with `uv sync --locked --extra train`. Preview with
-`uv run --extra train python examples/train_together.py`; append `--launch` to
+Install the locked SDK with `uv sync --locked`. Preview with
+`uv run python examples/train_together.py`; append `--launch` to
 upload and start a billed job. The script waits for both uploaded files to finish
 processing before submission. It uses the Python SDK directly to set LoRA rank.
 See Together's [fine-tuning API reference](https://docs.together.ai/reference/post-fine-tunes).
@@ -47,31 +47,31 @@ request before calling the endpoint. Inference and hosting have separate charges
 
 ## Follow the job and deploy from the CLI
 
-If using a `.env` file, load it explicitly for each command. Use the locked
-`train` extra so the Python example and `tg` use the same SDK:
+If using a `.env` file, load it explicitly for each command. The Together SDK and `tg` CLI are installed by `uv sync --locked`, so the
+Python example and CLI use the same version:
 
 ```bash
-uv sync --locked --extra train
+uv sync --locked
 cp .env.example .env
 # Edit .env to set TOGETHER_API_KEY, then preview or launch:
-uv run --extra train --env-file .env python examples/train_together.py
-uv run --extra train --env-file .env python examples/train_together.py --launch
+uv run --env-file .env python examples/train_together.py
+uv run --env-file .env python examples/train_together.py --launch
 ```
 
 Replace `YOUR_JOB_ID` below with the ID printed by that launch. Use the same ID
 for status and output-model lookup:
 
 ```bash
-uv run --extra train --env-file .env tg fine-tuning retrieve YOUR_JOB_ID
-uv run --extra train --env-file .env tg fine-tuning retrieve YOUR_JOB_ID --json
+uv run --env-file .env tg fine-tuning retrieve YOUR_JOB_ID
+uv run --env-file .env tg fine-tuning retrieve YOUR_JOB_ID --json
 ```
 
 Once training completes, copy `model_output_name` from the JSON. In the next
 command, replace `MODEL_OUTPUT_NAME` and choose hardware available to your account
 that supports your model. Creating the endpoint starts separately billed hosting:
 
 ```bash
-uv run --extra train --env-file .env tg endpoints create MODEL_OUTPUT_NAME \
+uv run --env-file .env tg endpoints create MODEL_OUTPUT_NAME \
   --hardware 1x_nvidia_h100_80gb_sxm --display-name tev1-4B-experimental --wait
 ```
 
@@ -89,7 +89,7 @@ model to emit that JSON object.
 Stop your dedicated endpoint when finished, using its **ID**:
 
 ```bash
-uv run --extra train --env-file .env tg endpoints stop YOUR_ENDPOINT_ID
+uv run --env-file .env tg endpoints stop YOUR_ENDPOINT_ID
 ```
 
 Training cost and duration vary with settings and provider capacity. The original

pyproject.toml

@@ -5,9 +5,4 @@ description = "Fine-tune Qwen3.5-4B decision models on Together AI"
 readme = "README.md"
 license = "MIT"
 requires-python = ">=3.12"
-dependencies = ["datasets==4.8.5", "transformers==5.17.0", "jinja2>=3.1,<4"]
-
-[project.optional-dependencies]
-train = [
-    "together==2.23.0",
-]
+dependencies = ["datasets==4.8.5", "transformers==5.17.0", "jinja2>=3.1,<4", "together==2.23.0"]

test_workflow.py

@@ -38,7 +38,7 @@ def test_sdk_serializes_completion_only_lora_recipe(self):
         try:
             from together import Together
         except ImportError:
-            self.skipTest('Install the train extra to verify SDK serialization')
+            self.fail('Run uv sync --locked to install the required Together SDK')
         client = Together(api_key='fixture')
         # Intercept submission and model-limits lookup so this test is offline.
         with patch.object(client.fine_tuning, 'estimate_price',