Aug 16 14:55 Name the proper-noun column for what it measures 14:52 Name the proper-noun column for what it measures 14:51 Name the proper-noun column for what it measures 14:42 Demote the proper-noun anecdote out of the result tables 14:41 Name the proper-noun column for what it measures 14:39 Name the proper-noun column for what it measures 14:33 v3.1 matrices, three-seed benchmark and eleven-language demo (part 2) 14:32 v3.1 matrices, three-seed benchmark and eleven-language demo Aug 14 18:02 Projection, seed and GPU model all act on the same free space 18:01 The plain matrices hold on a tightly written prompt 18:01 The plain matrices hold on a tightly written prompt 17:55 The plain matrices are not blind to what is stated once 17:55 The plain matrices are not blind to what is stated once 17:42 v3: calibrated against the stock 32B, image tokens for the 8B, demo folder Aug 12 06:46 Point people arriving from older tutorials at the current files 06:26 Point people arriving from older tutorials at the current files Aug 10 12:50 The -mlp residuals in fp16: half the download, and half the VRAM with node 0.1.4 12:21 README for 0.1.3: named people, the -mlp compatibility trap, corpus sources 12:19 ClipProj v0.1.3 matrices: 500 named people covered Ten matrices under mmh3-<encoder>-ClipProj[-celeb][-mlp]. The -celeb ones add 500 people to the calibration corpus: name tokens reconstruct at 0.8844 against 0.8265, for 0.007 of cosine on the general test set. The -mlp ones carry a residual network and need node 0.1.3, which an earlier node ignores silently. Previous matrices move to obsolete/ and the .pt copies are removed. Aug 9 17:42 Flatten the repo: ranking belongs in the README table, not the folder layout 17:42 Flatten the repo: ranking belongs in the README table, not the folder layout 17:42 Flatten the repo: ranking belongs in the README table, not the folder layout 17:42 Flatten the repo: ranking belongs in the README table, not the folder layout 17:41 Restore the .pt files alongside the safetensors until the reader has propagated 14:15 Correct the README: both formats are published during the transition 14:14 Restore the .pt files alongside the safetensors until the reader has propagated 14:12 Replaced by h3_qwen3vl_8b_tap24.safetensors: safetensors cannot execute code on load 14:12 Replaced by h3_qwen3vl_4b_int8convrot_tap24.safetensors: safetensors cannot execute code on load 14:12 Replaced by h3_qwen3vl_4b_tap24_200prompts.safetensors: safetensors cannot execute code on load 14:12 Replaced by h3_qwen3vl_4b_tap24.safetensors: safetensors cannot execute code on load 14:12 Replaced by h3_control_zero.safetensors: safetensors cannot execute code on load 14:12 Replaced by h3_control_identity.safetensors: safetensors cannot execute code on load 14:12 Replaced by h3_qwen3vl_8b_CONDPROJ_tap24.safetensors: safetensors cannot execute code on load 14:12 Replaced by h3_qwen3vl_4b_CONDPROJ_tap24.safetensors: safetensors cannot execute code on load 14:12 Promote the condition_proj-weighted matrices to the root, move the unweighted ones to obsolete/ 14:07 Moved to obsolete/h3_qwen3vl_8b_tap24.pt 14:07 Moved to obsolete/h3_qwen3vl_4b_int8convrot_tap24.pt 14:07 Moved to obsolete/h3_qwen3vl_4b_tap24_200prompts.pt 14:07 Moved to obsolete/h3_qwen3vl_4b_tap24.pt 14:07 Promote the condition_proj-weighted matrices to the root, move the unweighted ones to obsolete/ 14:05 Add the condition_proj-weighted matrices, calibrated against what the DiT actually reads 13:45 ClipProj v0.1.1: attention-sink vector added to every matrix, plus 8B and int8_convrot 09:42 ClipProj v0.1.0: projection matrices and controls 09:41 initial commit