mgoin/Qwen3-8B-speculator.dspark-reasoning

Qwen3-8B DSpark speculator (reasoning recipe)

DSpark draft model (DFlash backbone + Markov logit-bias head + confidence head) for Qwen/Qwen3-8B, trained with speculators (PR #677).

Trained to match the recipe of the reference DFlash speculator RedHatAI/Qwen3-8B-speculator.dflash: ultrachat answers regenerated by Qwen/Qwen3-8B with reasoning, sequence length 8192, lr 6e-4, cosine schedule, silu, max_anchors 3072, 3 epochs (50k samples).

DSpark adds a Markov logit-bias head (intra-block token dependency) and a per-position confidence head on top of the DFlash backbone; the matched-recipe DFlash above is the baseline reference.

Validation per-position acceptance (pos 1-7): 0.711 / 0.590 / 0.506 / 0.445 / 0.401 / 0.367 / 0.340 | full_acc 0.480 | confidence ECE-style abs error 0.054.