Draftfit · Humanizer 12B DFlash
A validation checkpoint for Draftfit — train DFlash-family drafts jointly, convert compatible families, and continue training existing drafts.
This DFlash draft accelerates Humanizer; it requires the target model.
Target / teacher: Humanizer. Initial draft weights: Gemma4 DFlash. This is a target-specific speculative draft, not a standalone Humanizer fine-tune.
Performance
| Model | tok/s | Acceptance | vs. original |
|---|---|---|---|
| Target only | 21.71 | — | — |
| Original DFlash | 23.43 | 30.1% | reference |
| Draftfit DSpark | 33.49 | 59.9% | +42.9% |
| Draftfit DFlash | 32.62 | 59.9% | +39.2% |
| Draftfit DFlash2 | 31.23 | 55.6% | +33.3% |
By domain · tok/s
| Model | General prose | Blog | Report | Essay | |
|---|---|---|---|---|---|
| Target only | 21.52 | 21.51 | 21.44 | 22.00 | 21.79 |
| Original DFlash | 25.22 | 24.38 | 23.75 | 22.01 | 22.98 |
| Draftfit DSpark | 33.59 | 39.29 | 33.54 | 31.77 | 30.93 |
| Draftfit DFlash | 32.92 | 38.81 | 32.80 | 29.75 | 29.13 |
| Draftfit DFlash2 | 33.16 | 35.81 | 30.46 | 31.14 | 29.48 |
Apple M5 Pro · MLX · mixed 4/8-bit target with FP32 floating computation · 4-bit drafts. Median throughput includes prefill and verification; acceptance is accepted/proposed draft tokens. Same 10 held-out inputs ×2 repeats per model, greedy, up to 64 output tokens, draft cap 3, lookup off. All 160 requests matched target-only tokens, including previous-checkpoint comparisons.
Continued drafts improved median throughput by 2.6–4.9% over their previous checkpoints in this run; trial variability is substantial, so this is not a robust or long-form speedup claim. Domain acceptance, previous checkpoints and details.
Acceptance by draft position
| Model | Position 1 | Position 2 | Position 3 |
|---|---|---|---|
| Original DFlash | 52.8% | 26.6% | 10.5% |
| Draftfit DSpark | 80.6% | 59.4% | 39.3% |
| Draftfit DFlash | 80.6% | 59.4% | 39.3% |
| Draftfit DFlash2 | 74.0% | 57.1% | 35.5% |
Same MLX runs above. Position k = accepted prefixes reaching k / rounds proposing position k. Earlier rejection also rejects later positions; these are not conditional token accuracies. Counts and domain breakdown.
How to use
MLX — tested. After runtime setup, pass prose to rewrite:
hf download jayyun98/Humanizer-12B-Draftfit-DFlash run_mlx.py --local-dir ./humanizer-draft
python ./humanizer-draft/run_mlx.py --text "The meeting has been moved to Friday." --check
vLLM / SGLang — unverified for this checkpoint. No CUDA serving or speed claim; use the tested MLX path.
Basic precision showed occasional token differences; the supplied runner uses the tested FP32 setting. Provenance · Apache-2.0 · DFlash MIT · Attribution.