From a39e806feced62b53318e6fa6707614bd7646179 Mon Sep 17 00:00:00 2001
From: Oasis-Git <ayw.sirius19@gmail.com>
Date: Tue, 13 Oct 2026 16:41:57 -0700
Subject: [PATCH] [DSv4.1] Add Flash model generation preset

---
 preset.md                   | 14 ++++++++++++++
 sgl_eval/model_presets.yaml |  7 +++++++
 tests/test_preset.py        | 21 ++++++++++++++++++++-
 3 files changed, 41 insertions(+), 1 deletion(-)

diff --git a/preset.md b/preset.md
index 30e2a28..12b096a 100644
--- a/preset.md
+++ b/preset.md
@@ -29,6 +29,20 @@ sgl-eval run aime25 \
   --load-preset-from-model-id deepseek-ai/DeepSeek-V4-Flash-0731
 ```
 
+### DeepSeek V4.1 Flash
+
+`deepseek-ai/DeepSeek-V4.1-Flash` uses temperature `1.0`, top-p `0.95`,
+thinking enabled, and reasoning effort `high`. It does not set `max_tokens`;
+the benchmark default applies (MMMU-Pro leaves the output limit to the server).
+For a bounded evaluation budget, pass `--max-tokens` explicitly.
+
+```bash
+sgl-eval run mmmu_pro \
+  --base-url http://localhost:30000/v1 \
+  --load-preset-from-model-id deepseek-ai/DeepSeek-V4.1-Flash \
+  --num-examples 200 --max-tokens 32768
+```
+
 Built-in model presets deliberately do not contain `base_url`; pass it on the
 CLI or provide it through a user preset. An unknown model ID exits before
 contacting the endpoint, prints the complete supported-ID list, and tells the
diff --git a/sgl_eval/model_presets.yaml b/sgl_eval/model_presets.yaml
index 12e2528..a3cb4d5 100644
--- a/sgl_eval/model_presets.yaml
+++ b/sgl_eval/model_presets.yaml
@@ -16,6 +16,13 @@ models:
       max_tokens: 400000
       thinking: true
       reasoning_effort: max
+  deepseek-ai/DeepSeek-V4.1-Flash:
+    model: deepseek-ai/DeepSeek-V4.1-Flash
+    sampling:
+      temperature: 1.0
+      top_p: 0.95
+      thinking: true
+      reasoning_effort: high
   moonshotai/Kimi-K3:
     model: moonshotai/Kimi-K3
     sampling:
diff --git a/tests/test_preset.py b/tests/test_preset.py
index cdf320b..dd0d875 100644
--- a/tests/test_preset.py
+++ b/tests/test_preset.py
@@ -7,7 +7,7 @@
 
 import pytest
 
-from sgl_eval.model_preset import ModelPreset
+from sgl_eval.model_preset import ModelPreset, load_model_preset
 from sgl_eval.preset import (
     Endpoint,
     Expected,
@@ -277,3 +277,22 @@ def test_reasoning_effort_validated_at_parse_time(raw, expected) -> None:
             reasoning_effort(raw)
     else:
         assert reasoning_effort(raw) == expected
+
+
+def test_dsv41_flash_model_preset() -> None:
+    preset = load_model_preset("deepseek-ai/DeepSeek-V4.1-Flash")
+    assert preset.model == "deepseek-ai/DeepSeek-V4.1-Flash"
+    gen = apply_to_gen(GenConfig(), None, _args(), preset)
+    assert gen.temperature == 1.0
+    assert gen.top_p == 0.95
+    assert gen.chat_template_kwargs == {"thinking": True}
+    assert gen.reasoning_effort == "high"
+    assert gen.max_tokens is None
+    assert gen.seed is None
+
+    overridden = apply_to_gen(
+        GenConfig(), None, _args(max_tokens=4096, thinking=False, reasoning_effort="low"), preset
+    )
+    assert overridden.max_tokens == 4096
+    assert overridden.chat_template_kwargs == {"thinking": False}
+    assert overridden.reasoning_effort == "low"