Collection No. 01 — 1,000 Tinker Recipes
Tinker Folio
One thousand fine-tuning recipes where the world's creative disciplines meet four Tinker kernels — instruction SFT, structured-output SFT, reward-based RL, and preference tuning — each one ready to ship in a single Lovable build.
Primary Directive · dance
Style Tag Weaver
Fine-tune Qwen/Qwen3-8B on 62 (clip, styleDSL) pairs so the model tags dance clips with valid style codes.
View Recipe
02
Shot List Rhyme
Theme · video
1kFine-Tune Recipes
Section IIFour Tinker kernels.
SFT-Instruct, SFT-Format, RL-Reward, RL-Preference — the four Tinker kernels every entry leans on.
Appendix · Build Strategy
→The Sandbox Training Protocol
Train once in Lovable's Linux sandbox. Ship the report page. Zero runtime API calls.
Status
Trained on Tinker.
Chapter I · Disciplines
See full index →Ten houses, one training loop.
💃01/10
Dance & Choreography
100 recipes →
🎵02/10
Music & Sound Design
100 recipes →
🎨03/10
Visual Art
100 recipes →
🎥04/10
Videography & Film
100 recipes →
📷05/10
Photography
100 recipes →
✍️06/10
Writing, Poetry & Narrative
100 recipes →
🎬07/10
Filmmaking & Animation
100 recipes →
🎮08/10
Game Design & Interactive Media
100 recipes →
🎭09/10
Theater & Live Performance
100 recipes →
👗10/10
Fashion & Textile Design
100 recipes →
Chapter II · Primitives
Read the primer →The four kernels.
supervised fine-tune№ 01
SFT · Instruction
Supervised fine-tuning on `(prompt, completion)` pairs. `service_client.create_lora_training_client(base_model=...)` → `forward_backward_async(data, 'cross_entropy')` → `optim_step_async(AdamParams(lr=1e-4))` → `save_weights_and_get_sampling_client(...)`. Teach a small open model to imitate a persona, a house style, or a domain of expertise.
format fine-tune№ 02
SFT · Structured Output
Supervised fine-tuning on `(input, structured_output)` pairs — JSON, DSLs, chord charts, poetic forms, lesson plans. Same Tinker SFT loop as instruction tuning, but the loss mask covers a rigid target schema so the model learns the shape as much as the content.
reinforcement learning№ 03
RL · Programmatic Reward
On-policy RL loop: `save_weights_and_get_sampling_client()` → `sample_async(num_samples=k)` rollouts → programmatic reward fn scores each rollout → `forward_backward_async(data, 'importance_sampling')` → `optim_step_async(...)` → repeat. Optimises the model for a task where you can code the grader.
taste fine-tune№ 04
RL · Preference Pairs
DPO-style preference tuning: dataset of `(prompt, chosen, rejected)` triples, `forward_backward_async` with a preference loss, `optim_step_async` steady LoRA. Teach the model taste — house voice, safety posture, tone — where good and bad examples are easy to label but a scalar reward is hard.
Three movements, one training run.
- Movement 0101Choose a recipeTen disciplines, four kernels each. Open the one that matches your team.
- Movement 0202Paste the mega-promptLovable's sandbox pip-installs tinker, runs the training script, writes the artifact.
- Movement 0303Ship the reportA single TanStack route renders the loss curve and before/after samples. No runtime API calls.
1,000 recipes indexed · zero runtime infra · ready to ship