Fireworks x HUD Release RL Training Cookbook: Define Task Once, Train and Evaluate LoRA
sophiamyang · x · 2026-09-26
Fireworks AI and HUD have released a cookbook showing how to train a LoRA adapter with reinforcement learning using HUD environments and the Fireworks Serverless Training API.
- Division of labor: HUD runs the rollouts — developers define task templates and programmatic graders there — while Fireworks handles the training infrastructure, converting rollout results into training datums and model updates.
- One definition, two uses: the same task definition serves both RL training and held-out evaluation.
- Details: each rollout returns a reward from the HUD grader plus a trace with the exact token IDs sampled by the model.
- Example: training a Qwen 3.8 27B LoRA adapter on four-digit multiplication in a local environment — a stand-in for any one-turn capability checkable by a programmatic grader.
More from Infra
- Training and Inference Share One GPU Fleet, but Fungibility Only Runs One Way — robleclerc · 2026-09-26
- Google Cloud brings native BM25 full-text search to AlloyDB and Cloud SQL via Tiger Data's pg_textsearch — michaelfreedman · 2026-09-26
- DuckDB-Wasm gets persistent browser databases via OPFS — with a version gotcha — JeremyCMorgan · 2026-09-26
- Goldman: AI Agent Token Use to Grow 24x by 2030 as Uber, Microsoft Rethink Costs — rohanpaul_ai · 2026-09-26
- Faster inference chips let you compress evals to test long-horizon tasks between model releases — gordic_aleksa · 2026-09-26
- Redditor wants to route coding tasks across 3 local machines, inspired by Claude Code's multi-model dispatch — Shadow_s_Bane · 2026-09-26