Enrich card: accuracy vs FLARE base + c=1 speed & commit-rate tables

README.md

@@ -12,28 +12,55 @@ tags:
 
 *English | [한국어](./README_ko.md)*
 
-Qwen3.5-4B FLARE two-stream **block-diffusion** checkpoint (research). Continued from the v6 SFT
-base (step_18000) with the **full-mask recipe**: 100% of each block masked (single view, no
+Qwen3.5-4B FLARE two-stream **block-diffusion** checkpoint (research). Continued from the FLARE-4B
+base (v6 SFT, step_18000) with the **full-mask recipe**: 100% of each block masked (single view, no
 complementary mask) at block size **bd=8**, fresh optimizer, LR 5e-5, ~2k steps.
 
 ## How to serve
 
 Best served **self-speculative** (draft with the diffusion stream, verify with the clean/AR stream):
-verify-guaranteed AR quality at ~2.6 accepted tokens/forward — the fastest self-spec of our
-checkpoints at concurrency 1. It also serves in plain **AR/causal** mode. **Pure iterative diffusion
-decoding is not supported by this recipe** (the shift decoder's partial-mask states are untrained).
+verify-guaranteed AR quality, and — because full-mask training shapes exactly the single-shot
+full-block prediction a self-spec draft needs — a **higher acceptance length than the base**, so it's
+the fastest self-spec at low concurrency. It also serves in plain **AR/causal** mode. **Pure iterative
+diffusion decoding is not supported by this recipe** (the shift decoder's partial-mask states are
+untrained). Runs on the HybridDiffusion SGLang fork. `mask_id = 248077`, vocab 248078.
 
-Runs on the HybridDiffusion SGLang fork. `mask_id = 248077`, vocab 248078.
+## Accuracy
 
-## Eval (concurrency 1, threshold 0.9)
+Concurrency 1, threshold 0.9, reasoning on. Compared against the **FLARE-4B base** (the v6 SFT
+checkpoint Trida2.0 continues from). *A vanilla Qwen3.5-4B baseline is being added.*
 
-| | self-spec | AR |
+| benchmark | Trida2.0 · self-spec | Trida2.0 · AR | FLARE base · self-spec |
+|---|---|---|---|
+| GSM8K | **84.2** | **89.3** | 83.2 |
+| MMLU-Pro | 63.0 | — | 63.5 |
+| IFEval (prompt-strict) | 69.7 | — | 70.1 |
+| GPQA-Diamond † | 36.4 | 38.4 | 40.9 |
+| BFCL (Non-Live AST) | **65.2** | — | 53.2 |
+| Ko-AgentBench L1 / L2 | **0.909 / 1.00** | — | 0.909 / 0.933 |
+| FunctionChat call / compl | 0.85 / 0.88 | — | 0.92 / 0.95 |
+| tau2 retail / air / tel | 0.40 / 0.50 / 0.80 | — | 0.65 / 0.55 / 0.90 |
+
+† GPQA-Diamond at a 16k token cap is heavily truncation-limited (long chain-of-thought); these are
+floors — a higher-cap rerun is pending.
+
+## Serving speed & parallel commit (concurrency 1)
+
+Block-diffusion self-spec commits **multiple tokens per forward pass** at verify-guaranteed AR
+quality. Trida2.0's full-mask training makes it a better drafter, so it accepts more per step and runs
+faster than the base.
+
+| model · decoding | tok/s (per-request) | tokens / forward |
 |---|---|---|
-| GSM8K | 84.2 | 89.3 |
-| MMLU-Pro | 63.0 | — |
-| IFEval (prompt-strict) | 69.7 | — |
-| BFCL (Non-Live AST) | 65.2 | — |
-| Ko-AgentBench L1/L2 | 0.909 / 1.00 | — |
-| FunctionChat call/compl | 0.85 / 0.88 | — |
+| **Trida2.0 · self-spec** | **243** | **2.59** |
+| Trida2.0 · AR | 216 | 1.00 |
+| FLARE base · self-spec | 188 | ~2.0 |
+| FLARE base · AR | 217 | 1.00 |
+| FLARE base · diffusion | 189 | ~2.5 |
+
+Trida2.0 self-spec is **~30% faster** than the base self-spec at equal-or-better accuracy. Per-benchmark
+commit rate (Trida2.0 self-spec, tokens/forward): **GSM8K 3.10, BFCL 2.33, MMLU-Pro 2.33**.
+
+*Concurrency-4 throughput and a vanilla Qwen3.5-4B baseline are being added.*
 
 Internal research checkpoint — trillionlabs.