Meta-Eval/EvalHarness
3
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Flatten config: direct key-value per bench, no nested groups

Browse Source 
Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
sora 2026-09-11 08:05:36 +00:00
parent d16009d054
commit 4d7567adf8
1 changed files with 42 additions and 110 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

152
evalharness/config/dp4-nothink.yaml
View File

@ -1,119 +1,51 @@
# dp4-nothink: DeepSeek-V4-Flash nothinking protocol
# Usage: evalharness eval run <bench> --config dp4-nothink --api-url ... --model ...
# Priority: DatasetSpec.gen_config < default < <bench> < CLI args
# dp4-nothink 配置
# 用法: evalharness eval run <bench> --config dp4-nothink --api-url ... --model ...
# 优先级: DatasetSpec < default < <bench> < 命令行参数
default:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0, stream: true}
run: {concurrency: 8, checkpoint: true, resume: true}
temperature: 0.0
max_tokens: 32768
top_p: 1.0
concurrency: 8
# ── Math (temp=1, 12 runs averaged) ──
aime24:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {repeats: 12, concurrency: 4}
# ── 数学（temp=1，跑 12 遍取均值）──
aime24: {temperature: 1.0, max_tokens: 8192, repeats: 12, concurrency: 4}
aime25: {temperature: 1.0, max_tokens: 8192, repeats: 12, concurrency: 4}
aime26: {temperature: 1.0, max_tokens: 8192, repeats: 12, concurrency: 4}
hmmt26: {temperature: 1.0, max_tokens: 8192, repeats: 12, concurrency: 4}
gsm8k: {limit: 200}
competition_math: {limit_per_task: 40}
aime25:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {repeats: 12, concurrency: 4}
# ── 知识/选择题 ──
mmlu: {limit_per_task: 4}
cmmlu: {limit_per_task: 3}
mmlu_pro: {limit_per_task: 20}
arc: {limit: 200}
hellaswag: {limit: 400}
winogrande: {limit: 200}
bbh: {limit_per_task: 10}
gpqa_diamond: {temperature: 1.0, max_tokens: 8192, limit: 200}
aime26:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {repeats: 12, concurrency: 4}
# ── 问答 ──
simple_qa: {max_tokens: 1024, limit: 200}
trivia_qa: {limit: 200}
drop: {limit: 200}
hmmt26:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {repeats: 12, concurrency: 4}
# ── Judge 类 ──
hle: {limit_per_task: 25, judge: dp4-flash, judge_url: 'http://174.1.51.4:30000/v1'}
imo_answerbench: {temperature: 1.0, max_tokens: 8192, limit_per_task: 25, judge: dp4-flash, judge_url: 'http://174.1.51.4:30000/v1'}
competition_math:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0}
run: {limit_per_task: 40, concurrency: 8}
# ── 代码（docker 沙箱）──
humaneval: {temperature: 1.0, concurrency: 4}
bigcodebench: {limit: 200, concurrency: 4}
live_code_bench: {temperature: 1.0, limit: 200, concurrency: 4}
gsm8k:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0}
run: {limit: 200, concurrency: 8}
# ── 长上下文 ──
longbench_v2: {max_tokens: 8192, limit_per_task: 66, concurrency: 4, max_input_tokens: 120000}
openai_mrcr: {max_tokens: 8192, limit: 200, concurrency: 4, max_input_tokens: 120000}
# ── Knowledge / MCQ ──
mmlu:
run: {limit_per_task: 4, concurrency: 8}
cmmlu:
run: {limit_per_task: 3, concurrency: 8}
mmlu_pro:
run: {limit_per_task: 20, concurrency: 8}
arc:
run: {limit: 200, concurrency: 8}
hellaswag:
run: {limit: 400, concurrency: 8}
winogrande:
run: {limit: 200, concurrency: 8}
bbh:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0}
run: {limit_per_task: 10, concurrency: 8}
gpqa_diamond:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {limit: 200, concurrency: 8}
# ── QA ──
simple_qa:
generation: {temperature: 0.0, max_tokens: 1024, top_p: 1.0}
run: {limit: 200, concurrency: 8}
trivia_qa:
run: {limit: 200, concurrency: 8}
drop:
run: {limit: 200, concurrency: 8}
# ── Judged ──
hle:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0}
run: {limit_per_task: 25, concurrency: 8}
judge: {model: dp4-flash, api_url: 'http://174.1.51.4:30000/v1'}
imo_answerbench:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {limit_per_task: 25, concurrency: 8}
judge: {model: dp4-flash, api_url: 'http://174.1.51.4:30000/v1'}
# ── Code (docker sandbox) ──
humaneval:
generation: {temperature: 1.0, max_tokens: 32768, top_p: 1.0}
run: {concurrency: 4}
bigcodebench:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0}
run: {limit: 200, concurrency: 4}
live_code_bench:
generation: {temperature: 1.0, max_tokens: 32768, top_p: 1.0}
run: {limit: 200, concurrency: 4}
# ── Long context ──
longbench_v2:
generation: {temperature: 0.0, max_tokens: 8192, top_p: 1.0}
run: {limit_per_task: 66, concurrency: 4, max_input_tokens: 120000}
openai_mrcr:
generation: {temperature: 0.0, max_tokens: 8192, top_p: 1.0}
run: {limit: 200, concurrency: 4, max_input_tokens: 120000}
# ── Agent / tools ──
bfcl_v3:
generation: {temperature: 0.0, max_tokens: 4096, top_p: 1.0, parallel_tool_calls: true}
run: {limit_per_task: 20, concurrency: 8, env: bfcl_mock}
general_fc:
generation: {temperature: 0.0, max_tokens: 4096, top_p: 1.0}
run: {limit: 400, concurrency: 8}
tau2_bench:
generation: {temperature: 0.0, max_tokens: 16384, top_p: 1.0}
run: {concurrency: 8, env: tau2_official, max_turns: 50}
swe_bench_verified:
run: {limit: 70, concurrency: 4}
# ── Agent ──
bfcl_v3: {max_tokens: 4096, limit_per_task: 20, env: bfcl_mock}
general_fc: {max_tokens: 4096, limit: 400}
tau2_bench: {max_tokens: 16384, env: tau2_official, max_turns: 50}
swe_bench_verified: {limit: 70, concurrency: 4}