Releases: p-e-w/heretic
Releases · p-e-w/heretic
v1.4.0
6ea3b8d
This commit was signed with the committer’s verified signature.
p-e-w
Philipp Emanuel Weidmann
Changes
- @p-e-w implemented automatically reproducing a model from a
reproduce.jsonfile in #326. @Vinay-Umrethe fixed issues in that implementation in #352. - @rocker-zhang added support for plain text files as prompt datasets in #337, based on earlier work by @ricyoung in #103.
- @anrp fixed LoRA export in #321
- @zaakirio implemented saving the processor for multimodal models in #353
- @MoonRide303 added support for gemma-4-12B-it in #350
- @coder3101 added support for LiquidAI/LFM2.5 in #344
- @UnstableLlama add a config file for suppressing humor in #340
- @umran666 fixed a trial counting bug in #357
- @umran666 made resetting the model null-safe to handle study cancellations in #367
- @umran666 improved exception formatting in #363
- @umran666 resolved a variable shadowing issue in #356
- @iuyua9 fixed recognizing root Hugging Face repo IDs in #325
- @rocker-zhang removed the unused
hf-transferdependency in #338 - @kabachuha added "disclaimer" to the prohibited strings list in #334
- @Vinay-Umrethe removed the annoying gray background from the README logo in #359
- @p-e-w added functionality for collecting
reproduce.jsonfiles from Hugging Face, fixed akernelsdependency problem, as well as other minor issues
New Contributors
- @iuyua9 made their first contribution in #325
- @kabachuha made their first contribution in #334
- @rocker-zhang made their first contribution in #337
- @UnstableLlama made their first contribution in #340
- @coder3101 made their first contribution in #344
- @zaakirio made their first contribution in #353
- @umran666 made their first contribution in #356
Full Changelog: v1.3.0...v1.4.0
Contributors
ricyoung, anrp, and 10 other contributors
Assets 4
v1.3.0
9b7624d
This commit was signed with the committer’s verified signature.
p-e-w
Philipp Emanuel Weidmann
Changes
- @Vinay-Umrethe (who had previously contributed under the username @Vinayyyy7) implemented reproducible runs in #191. @p-e-w revised and improved that implementation in #303.
- @magiccodingman reduced peak VRAM usage in #239. @olekssy fixed a bug in that implementation in #301.
- @farolone added support for Qwen3.5 models in #187
- @MoonRide303 added support for Gemma 4 models in #287
- @erm14254 made sure all abliterable components across layers are displayed in #215
- @cpagac fixed VRAM usage reporting for multi-GPU setups in #169
- @cpagac fixed a division-by-zero error in the evaluator in #225
- @spikymoth improved automatic response prefix determination with a two-step process in #194
- @spikymoth added model card generation for local models with an existing README in #157
- @Diplo2by improved startup speed when Heretic is run with
-h/--helpin #293 - @AWuhrmann fixed the example value for the
max_memorysetting in #284 - @p-e-w added an integrated benchmarking system, made the response prefix logic configurable, implemented multiple infrastructure improvements, and fixed various minor issues
New Contributors
- @cpagac made their first contribution in #169
- @farolone made their first contribution in #187
- @erm14254 made their first contribution in #215
- @AWuhrmann made their first contribution in #284
- @MoonRide303 made their first contribution in #287
- @Diplo2by made their first contribution in #293
- @magiccodingman made their first contribution in #239
- @olekssy made their first contribution in #301
Full Changelog: v1.2.0...v1.3.0
Contributors
farolone, p-e-w, and 9 other contributors
Assets 4
v1.2.0
27097bf
This commit was signed with the committer’s verified signature.
p-e-w
Philipp Emanuel Weidmann
Changes
- @noctrex added a
max_memorysetting to limit memory usage in #83 - @spikymoth added a mechanism to avoid excessive low-divergence iteration in #73
- @accemlcc implemented a new LoRA-based abliteration engine with support for 4-bit quantization in #60
- @accemlcc added enumeration of all available GPUs on startup in #86
- @Vinayyyy7 added the ability to run more trials after optimization is complete in #76
- @anrp fixed MXFP4 loading in #107
- @anrp refactored the save machinery in #110
- @anrp added broad support for VL models in #108
- @anrp implemented saving and resuming optimization progress in #106, #119, and #116
- @spikymoth implemented Magnitude-Preserving Orthogonal Ablation in #52
- @salmanmkc upgraded GitHub Actions to the latest versions in #136 and #137
- @p-e-w added full type checking of the codebase, debug output, prompt modification functionality, and an example config file for slop reduction, and fixed various minor issues
New Contributors
- @noctrex made their first contribution in #83
- @accemlcc made their first contribution in #60
- @anrp made their first contribution in #107
- @salmanmkc made their first contribution in #136
Full Changelog: v1.1.0...v1.2.0
Contributors
anrp, p-e-w, and 4 other contributors
Assets 4
v1.1.0
d9f2b04
This commit was signed with the committer’s verified signature.
p-e-w
Philipp Emanuel Weidmann
Changes
- @mbarnson added basic MPS (Apple Silicon) support in #5
- @red40maxxer reduced memory usage in #15
- @Ooooze added IBM Granite MoE support in #14
- @kldzj added multi-GPU support in #17 and #32
- @ricyoung fixed an error when Hugging Face user profile fields are missing in #20
- @tymat added support for MXFP4 quantized models with Triton tensors in #28
- @spikymoth improved support for loading local datasets in #33
- @kldzj added support for models that require
trust_remote_codein #31 - @Vinayyyy7 added notebook (Colab/Kaggle) compatibility in #42
- @Vinayyyy7 fixed loading for certain models that default to the float32 dtype in #44
- @spikymoth improved refusal detection in #45
- @red40maxxer added a PR title lint to CI in #66
- @p-e-w added research features, support for stopping the optimization process early, and support for thinking models, and implemented an important padding fix suggested by @accemlcc
New Contributors
- @mbarnson made their first contribution in #5
- @red40maxxer made their first contribution in #15
- @Ooooze made their first contribution in #14
- @kldzj made their first contribution in #17
- @ricyoung made their first contribution in #20
- @tymat made their first contribution in #28
- @spikymoth made their first contribution in #33
- @Vinayyyy7 made their first contribution in #42
Full Changelog: v1.0.1...v1.1.0
Contributors
ricyoung, tymat, and 7 other contributors