| --- |
| license: apache-2.0 |
| language: |
| - it |
| library_name: echo-1.58 |
| tags: |
| - masked-diffusion |
| - ternary |
| - 1.58-bit |
| - quantization-aware-training |
| - research-checkpoints |
| pipeline_tag: fill-mask |
| --- |
| |
| # Echo-1.58 27M Research Ladder |
|
|
| A bundle of 27M-parameter checkpoints (`d_model` 448, 8 layers) that back the paper's |
| mechanistic results: the recovery ablation, the multi-epoch repetition ladder (1 to 20 |
| epochs on a fixed 75M-token pool), the technique arms, and the fresh-data baselines. Each |
| file is bfloat16, paired with its FP16 twin where relevant. These are provenance artifacts for |
| reproducing the tables and figures, not standalone models. See |
| https://github.com/lupodevelop/echo-1.58 for the code that consumes them and |
| `results/RESULTS.md` for the numbers. |
|
|
| ## Contents |
|
|
| - `fp16_e{1,5,10,20}.pt`, `tern_e{5,10,20}.pt`, `rec_e{5,10}.pt` repetition ladder |
| - `rec_e5_lr3.pt`, `tern_e5_lr3.pt`, `rec_e5_vsched.pt`, `rec_e5_lr8.pt`, `rec_e5_rewarm.pt` |
| technique and learning-rate arms |
|
|
