Render tool calls properly, follow the Hub's theme, fix the model link

Handle a dataset the viewer hasn't finished indexing

Correct the speedup figures to within-process measurements

Document the second optimization pass in RECIPES.md

Choose max_length per dataset and turn off gradient checkpointing

Support tool-calling datasets in preflight and preview

Say plainly that the 16-example batch target mostly does not bind

Point the recipe's own default and notes at 32768

Raise max_length to 32768 for every model

Show what a run will cost before the user starts it

Adopt LR 1e-3 and lora_alpha 32, and retract the confounded rank sweep

Cap rank at 64, epochs at 2, batch target at 16 — from a 2026 SFT sweep

Stop a diverged run from pushing a broken model

One flavor for every model, not 'all three sizes' — there are six recipes now

Fix comments that ended up describing the wrong recipe entries

Tidy a stale docstring: TOKENS_PER_STEP is computed per model now

Long-conversation data exposed two bugs; fix both

Add Phi-3.5-mini as the fourth family, and record why a fourth was hard

Make the push replace the repo contents instead of accumulating

Liger timings for all five models; gpt-oss becomes the cheapest of the big ones

Enable the Liger kernel: 31% faster on both dense models, for nothing

Push the adapter for gpt-oss: merging it produces a broken model

Add SmolLM3 and gpt-oss recipes, measured; restructure for four families

Derive the job timeout from the estimate — the fixed table was silently fatal

Add jinja2 — apply_chat_template needs it and transformers doesn't declare it

Train at 16384, and add the preflight route that was missing

Measure tokens instead of assuming them, and train at 8192 instead of 2048

Label the flavor table as SDPA and carry the FA3 number alongside it

Use FlashAttention 3, which reverses the batch-shape finding

Load FlashAttention via kernels, and scope the batch-shape claim to SDPA

Record why lora_alpha stays fixed as the rank moves

Order the before/after by how much actually changed

Add huggingface_hub to the server's requirements, and a demo dataset with a voice

Make it real: submit actual training jobs, read everything from the Hub

Cut the API back to one endpoint

Add the HTTP API, with execution faked

Pin server deps to versions that exist

Finetune Studio: no-code fine-tuning UI, fully mocked