Dream-7B-Instruct-s1k-sft

Dream-7B-Instruct-s1k-sft is a diffusion-based instruct model post-trained from Dream-v0-Instruct-7B on simplescaling/s1K, using MDLM (masked diffusion) and trained with the dLLM framework.

Model Overview

Dream-7B-Instruct-s1k-sft has the following features:

• Method: Masked Diffusion Language Modeling (MDLM);
• Framework: dLLM
• Base model: Dream-v0-Instruct-7B
• Dataset (SFT): simplescaling/s1K

For broader training and ablation reporting in the dLLM ecosystem, see the dLLM paper.

Eval notes: Metrics use confidence-threshold decoding (alg: confidence_threshold). The primary table is at confidence_threshold = 0.9; full grids sweep confidence_threshold ∈ {0.6, 0.7, 0.8, 0.9} with max_new_tokens ∈ {256, 512}.

---

Primary results

Benchmark	max_new_tokens=256 (Acc % | TPS)	max_new_tokens=512 (Acc % | TPS)
GSM8K	81.80 | 2.30	84.31 | 2.56
HumanEval	54.27 | 2.53	53.05 | 3.24
MBPP	57.80 | 2.21	57.80 | 2.37
MATH	45.16 | 2.32	49.70 | 2.57

---

Threshold sweep

Benchmark	τ=0.6 Acc | TPS	τ=0.7 Acc | TPS	τ=0.8 Acc | TPS	τ=0.9 Acc | TPS
GSM8K	65.66 | 4.14	74.68 | 3.38	79.83 | 2.66	81.80 | 2.30
HumanEval	34.15 | 4.01	43.90 | 3.53	51.83 | 3.00	54.27 | 2.53
MBPP	41.80 | 3.53	49.40 | 2.71	55.80 | 2.36	57.80 | 2.21
MATH	37.02 | 3.64	41.80 | 3.12	44.48 | 2.67	45.16 | 2.32

max_new_tokens=256, columns sweep confidence_threshold ∈ {0.6, 0.7, 0.8, 0.9}

Benchmark	τ=0.6 Acc | TPS	τ=0.7 Acc | TPS	τ=0.8 Acc | TPS	τ=0.9 Acc | TPS
GSM8K	67.55 | 6.45	75.44 | 5.34	81.65 | 3.54	84.31 | 2.56
HumanEval	31.10 | 4.42	45.73 | 3.98	50.00 | 3.54	53.05 | 3.24
MBPP	42.20 | 5.60	50.20 | 3.20	56.80 | 2.59	57.80 | 2.37
MATH	38.72 | 4.68	44.10 | 3.90	48.22 | 3.13	49.70 | 2.57

max_new_tokens=512, columns sweep confidence_threshold ∈ {0.6, 0.7, 0.8, 0.9}