Ask the viewer what a dataset contains, and show fewer of them dec9131 qgallouedec HF Staff commited on Jul 27
Fix dataset search and the runs list; add model logos; reclaim vertical space 0a5812a qgallouedec HF Staff commited on Jul 27
Render tool calls properly, follow the Hub's theme, fix the model link 496db9b qgallouedec HF Staff commited on Jul 26
Tidy OPTIMIZATION.md: fold the in-flight section into results 0289db8 qgallouedec HF Staff commited on Jul 26
Record the second pass results and the one remaining lever f3f3edd qgallouedec HF Staff commited on Jul 26
Correct the speedup figures to within-process measurements f5a74c4 qgallouedec HF Staff commited on Jul 26
Choose max_length per dataset and turn off gradient checkpointing 8ae7372 qgallouedec HF Staff commited on Jul 26
Support tool-calling datasets in preflight and preview 5849796 qgallouedec HF Staff commited on Jul 26
Say plainly that the 16-example batch target mostly does not bind 0fefead qgallouedec HF Staff commited on Jul 26
Adopt LR 1e-3 and lora_alpha 32, and retract the confounded rank sweep 2b445be qgallouedec HF Staff commited on Jul 26
Cap rank at 64, epochs at 2, batch target at 16 — from a 2026 SFT sweep d7a4218 qgallouedec HF Staff commited on Jul 26
Cross-reference the upstream writeup for the NaN failure 3db8df4 qgallouedec HF Staff commited on Jul 26
Check the new filter doesn't over-drop, which would silently shrink training 6a87007 qgallouedec HF Staff commited on Jul 26
Record what was actually validated end to end, not just benchmarked e955dc3 qgallouedec HF Staff commited on Jul 26
One flavor for every model, not 'all three sizes' — there are six recipes now b7720aa qgallouedec HF Staff commited on Jul 26
Fix comments that ended up describing the wrong recipe entries efc6bd2 qgallouedec HF Staff commited on Jul 26
Tidy a stale docstring: TOKENS_PER_STEP is computed per model now 93f59e6 qgallouedec HF Staff commited on Jul 26
README: four families, and surface the learning-rate cliff up front ffd4d85 qgallouedec HF Staff commited on Jul 26
Add Phi-3.5-mini as the fourth family, and record why a fourth was hard 7c7238a qgallouedec HF Staff commited on Jul 26
Make the push replace the repo contents instead of accumulating 9a65f81 qgallouedec HF Staff commited on Jul 26
Write up the MXFP4 merge bug that produced a broken gpt-oss model 1296165 qgallouedec HF Staff commited on Jul 26
Gemma 4's verdict: it works, comes last, and every penalty is TRL's VLM gate a1fef2b qgallouedec HF Staff commited on Jul 26
Record the learning-rate cliff: below 1e-4 nothing transfers ebe0603 qgallouedec HF Staff commited on Jul 26
Liger timings for all five models; gpt-oss becomes the cheapest of the big ones c3c21ba qgallouedec HF Staff commited on Jul 26
Enable the Liger kernel: 31% faster on both dense models, for nothing e7c7f30 qgallouedec HF Staff commited on Jul 26
Push the adapter for gpt-oss: merging it produces a broken model 0b2f0d5 qgallouedec HF Staff commited on Jul 26
Move the recipe detail to RECIPES.md and correct a stale status table 1eac071 qgallouedec HF Staff commited on Jul 26
Offer three families in the UI, and stop claiming bigger means slower 01f9627 qgallouedec HF Staff commited on Jul 26
Add SmolLM3 and gpt-oss recipes, measured; restructure for four families 504bb28 qgallouedec HF Staff commited on Jul 26
Derive the job timeout from the estimate — the fixed table was silently fatal 695da39 qgallouedec HF Staff commited on Jul 26
Bring the docs and contract in line with what actually ships cdf8cc3 qgallouedec HF Staff commited on Jul 26
Add jinja2 — apply_chat_template needs it and transformers doesn't declare it 17edcea qgallouedec HF Staff commited on Jul 26
Don't lead with an example whose viewer index is still building ebdc9f3 qgallouedec HF Staff commited on Jul 26
Train at 16384, and add the preflight route that was missing 4b03457 qgallouedec HF Staff commited on Jul 26
Measure tokens instead of assuming them, and train at 8192 instead of 2048 a8f0000 qgallouedec HF Staff commited on Jul 26
Label the flavor table as SDPA and carry the FA3 number alongside it acada3f qgallouedec HF Staff commited on Jul 26
Use FlashAttention 3, which reverses the batch-shape finding 4b66ae2 qgallouedec HF Staff commited on Jul 26
Load FlashAttention via kernels, and scope the batch-shape claim to SDPA 9205960 qgallouedec HF Staff commited on Jul 26
qgallouedec HF Staff commited on 