EvalHarness/evalharness/config/dp4-nothink.yaml
2026-09-11 08:02:43 +00:00

120 lines
3.3 KiB
YAML

# 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
default:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0, stream: true}
run: {concurrency: 8, checkpoint: true, resume: true}
# ── Math (temp=1, 12 runs averaged) ──
aime24:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {repeats: 12, concurrency: 4}
aime25:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {repeats: 12, concurrency: 4}
aime26:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {repeats: 12, concurrency: 4}
hmmt26:
generation: {temperature: 1.0, max_tokens: 8192, top_p: 1.0}
run: {repeats: 12, concurrency: 4}
competition_math:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0}
run: {limit_per_task: 40, concurrency: 8}
gsm8k:
generation: {temperature: 0.0, max_tokens: 32768, top_p: 1.0}
run: {limit: 200, concurrency: 8}
# ── 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}