fw-ai/forge-workshops

fw-ai/forge-workshops

Description: Fireworks Forge Workshops

Language: Jupyter Notebook

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-10-23T15:55:24Z

Pushed: 2026-10-26T20:38:28Z

Default branch: main

Fork: no

Archived: no

README:

Forge workshops

Hands-on workshops for training and using models with Fireworks. Each workshop lives in its own folder with its own notebooks, requirements and README, and needs only a Fireworks API key.

| Workshop | What you build | Time | Slides | |---|---|---|---| | [research-agent-rl/](research-agent-rl/) | Train a multi-hop research agent (Kimi K3) with reinforcement learning, then run it as a LangChain agent | ~20 min train, ~20-30 min LangChain (includes deploying the trained model) | Research Agent | | [decision-router/](decision-router/) | Train a model router (Qwen 3.8) with the cookbook's decision loop, then route a LangChain agent between a cheap and a strong model | ~15-20 min train, ~5-10 min LangChain | Decision Model | | [skill-to-weights/](skill-to-weights/) | Mini workshop: train a SKILL.md into GLM 5.3 Flash's weights with one short SFT run, scored by a plain Python function | ~15 min | - |

All workshops use the Fireworks Training SDK (fireworks-ai 1.2.27) and install the Fireworks training cookbook as a pip dependency, so there is nothing to clone.

Getting started

Use one virtual environment per workshop: their pinned package versions differ.

cd decision-router # or: cd research-agent-rl, cd skill-to-weights
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt # do this before the session; the first install downloads PyTorch
cp .env.example .env # set FIREWORKS_API_KEY
jupyter lab

Then open the workshop's workshop_train_*.ipynb (or workshop_skill_to_weights.ipynb) and run it top to bottom. Each workshop's README lists its notebooks, expected time per stage, files, and anything account-specific.

Layout

forge-workshops/
README.md this index
.gitignore .env, venvs, run outputs, downloaded data
research-agent-rl/
README.md requirements.txt .env.example
workshop_train_research_agent.ipynb workshop_langchain_research_agent.ipynb
research_env.py research_rollout.py
data/ MuSiQue downloads here on first run (not committed)
runs/ closed-book caches (committed); per-run logs (not committed)
decision-router/
README.md requirements.txt .env.example
workshop_train_router.ipynb workshop_langchain_router.ipynb
runs/ cached replay answers (committed); per-run rows, logs, charts (not committed)
skill-to-weights/
README.md requirements.txt .env.example
workshop_skill_to_weights.ipynb
skills/caveman/ vendored skill + license
cases/caveman.jsonl 40 cases with grading labels
runs/ per-run replies and result.png (not committed)