| # Reproduction Matrix |
|
|
| This matrix is the execution checklist for source-code reproduction first, then |
| basecode reproduction with matched metrics. |
|
|
| | Paper / route | Source code path | Required assets | Source metrics to reproduce | Basecode hook | Status | |
| | --- | --- | --- | --- | --- | --- | |
| | LLaDA/Dream-style block diffusion | `reference/code/Nemotron-Labs-Diffusion`, `reference/code/mbd-lms` | `llada2-mini`, `dream-v0-base-7b`, GSM8K/MATH/HumanEval/MBPP | accuracy, NFE, latency, tokens per forward | `ConfidenceRemaskSampler`, `MultiBlockSampler` | planned | |
| | MBD-LMs | `reference/code/mbd-lms` | `llada2-mini`, math/code datasets | multi-block speed/quality tradeoff | `MultiBlockSampler` | planned | |
| | DMax | `reference/code/DMax` | `dmax-16b`, GSM8K/MATH/code evals | aggressive parallel decoding accuracy and speed | `DMaxSampler` | planned | |
| | TAD / trajectory self-distillation | `reference/code/TAD`, `reference/code/T3D` | trajectory datasets, distilled checkpoints | few-step quality retention | `DMaxSampler` with distilled adapter | planned | |
| | Fast-dLLM | `reference/code/Fast-dLLM` | discrete dLLM checkpoint | KV-cache speedup, parallel decode latency | adapter cache API | planned | |
| | DFlash / PRESTO | `reference/code/dflash`, Nemotron references | draft model + AR target | acceptance length, speedup, target quality | `SpeculativeSampler` | planned | |
| | dMoE | `reference/code/dMoE` | `dmoe-16b` | expert/block routing quality and speed | expert adapter extension | planned | |
| | ELF / continuous flows | `reference/code/ELF` | ELF checkpoints, OpenWebText/LM1B | perplexity, MAUVE/diversity, generation speed | separate continuous adapter | planned | |
| | FLM/FMLM one-step | `reference/code/flm` | manual checkpoints, OpenWebText/LM1B | unconditional generation PPL/time | separate continuous adapter | planned | |
|
|
| Run order: |
|
|
| 1. Confirm remote assets with `python3 scripts/download_assets.py --dry-run`. |
| 2. Download `core` tier only. |
| 3. Run each source repository's smallest documented eval. |
| 4. Normalize metrics into `runs/source_reproductions/<paper_id>/metrics.json`. |
| 5. Implement/attach matching adapter in `basecode/blockdiff/adapters/`. |
| 6. Run `basecode` evals into `runs/basecode/<method_id>/metrics.json`. |
| 7. Compare source vs basecode metrics in one table. |
|
|
|
|
