Post-training

You're Claude Code, working in my repository.

Task
Choose the post-training recipe that fits our goal and GPUs (TorchTitan SFT or RL for Muse Spark Open, or Muse Glimmer's Axolotl, TRL or TitanRL paths), tell me which and why, and set up its first run in this project: the install, the smoke-test command, and a check of our data in its row format.

Read this first
Meta's docs page "Post-training" as Markdown: https://dev.meta.ai/docs/cookbook/post-training.md
It's the source of truth for request and response shapes. The same page for people: https://dev.meta.ai/docs/cookbook/post-training

Setup
- Server: the model runs on hardware you control, behind an OpenAI-compatible server (SGLang, vLLM, llama.cpp or another runtime this page covers). Code takes the server's base URL from configuration (such as http://127.0.0.1:8000/v1) and sends the model name it serves (GET <base URL>/models lists it).
- Auth: a local server takes any placeholder key unless it was started with one; read a real key, and any token a download needs, from the environment. Never hard-code, print or commit a key or token.
- Python: the OpenAI SDK (pip install openai) with base_url set to the server's /v1 and a placeholder api_key, or plain requests.

Facts to get right
- TorchTitan SFT: a one-GPU debug smoke test, then 128 GPUs for the 816B recipe; rows are {"messages": [...]} with image and audio parts by path.
- TorchTitan RL: GRPO with the DAPO loss; the published recipe uses 256 trainer GPUs plus 16 generator GPUs, and a five-step bring-up.
- Muse Glimmer: Axolotl on 8 × H100 (QLoRA also on 1 × H100); TitanRL's data-analyst recipe on one 8 × H100 node in about 30 minutes for the smoke config.

Starter code (Python)
import json

for i, line in enumerate(open("train.jsonl"), 1):
    if line.strip():
        row = json.loads(line)
        assert any(m["role"] == "assistant" for m in row["messages"]), f"line {i}: nothing to learn"

Done when
- The smoke test runs and logs a step-1 loss (near 12.2 for TorchTitan's SFT debug model).
- The data check passes on our rows.
- Any key or token comes from the environment and never appears in code, logs or commits.