Trim README: drop training details, condense architecture and acceptance notes

README.md

@@ -6,28 +6,21 @@ private: true
 # Muse-Spark-1.2-DSpark
 
 DSpark speculative-decoding draft model for **Muse-Spark-1.2-816B-A42B-NVFP4-open**
-(816B-A42B MoE target, NVFP4). Current head: **fresh_fp32_20260920 final, step 3600**
-(commit `6f102eebc1aa`).
+(816B-A42B MoE target, NVFP4).
 
 ## DSpark architecture
 
-- 5-layer decoder-only draft: hidden 8192, 128 attention heads / 16 KV heads, FFN 16384, BF16
-- Causal sliding-window attention (window 2048); RoPE theta 500,000; positions up to 1M
-- Reads target hidden states tapped at layers [13, 29, 45, 57, 61] (of 62) and reuses the
-  last target hidden state as an auxiliary input
-- Full target vocabulary (202,048 tokens) with the frozen target LM head; identity `d2t` mapping
-- Proposes **7 tokens per step** (`block_size=7`, sample-from-anchor); learned confidence head
-  with Markov head (rank 512)
-- Trained from scratch (FP32 master weights, 3,600 AdamW updates, LR 3e-4 cosine) on on-policy
-  data regenerated by the target; loss 0.1·CE + 1.8·TV with positional decay γ=4
+- 5-layer decoder-only draft (hidden 8192, 128/16 attention/KV heads, FFN 16384), BF16,
+  causal sliding-window attention (2048)
+- Full 202,048-token target vocabulary; frozen target LM head
+- Conditioned on target hidden states tapped at layers [13, 29, 45, 57, 61] plus the
+  final hidden state
+- Proposes 7 tokens per step (sample-from-anchor); learned Markov confidence head (rank 512)
 - ~10.6 GB, 74 tensors, safetensors
 
-## Acceptance rates (full suite, block verification)
+## Acceptance rates
 
-Served with the Muse Spark mpk runtime, TP4/EP, BF16, eager, no prefix caching;
-7 probabilistic draft tokens, adaptive verification off; acceptance arms at temperature 1.0 /
-top-p 0.95 (MQA at 0.7). Acceptance = accepted / proposed draft tokens; tokens/cycle includes
-the target token; positions divide by all verification cycles.
+Measured with 7 probabilistic draft tokens; temperature 1.0 / top-p 0.95 (MQA at 0.7).
 
 | Dataset | Requests | Acceptance | Tokens/cycle | Positions 1–7 (%) |
 |---|---:|---:|---:|---|
@@ -53,4 +46,4 @@ GSM8K MQA accuracy (separate from acceptance): **94.01%** (1,319 requests, 5-sho
  "rejection_sample_method":"block"}
 ```
 
-Requires the Muse Spark mpk serving runtime (`zixi/muse-spark-hf-refresh-20260916`).
+Requires a Muse Spark–compatible serving runtime.