| language: en | |
| license: apache-2.0 | |
| tags: | |
| - diffusion-language-model | |
| - masked-diffusion | |
| - mamba | |
| - experimental | |
| # d1-135m-28b (development archive) | |
| This is the development archive of the DIMBA project: every checkpoint from the full chronology, including the failed continuous-diffusion first attempt, the EOS-collapse SFT, and other negative results. Kept public because negative results are results. | |
| **The final release lives at [devnull37/hr-diffuse-1-nano](https://huggingface.co/devnull37/hr-diffuse-1-nano)**, with the release model, the repair variant, the token-critic head, and the full system card. Start there. | |
| Checkpoint chronology in this repo: `kd_final.pt` (continuous KD, negative), `mdm_final.pt` (first masked-diffusion base), `mdm2_final.pt` (fresh-data base, capacity-bound negative), `mdm_sft_final.pt` (EOS-collapse failure), `mdm_sft2_final.pt`, `mdm2_sft_final.pt`, `mdm_sft_cfg_final.pt` (first CFG model), `mdm_sft_cfg2_final.pt` (release candidate, published in the release repo as `hr_diffuse_1_nano.pt`). | |