PyPI - local-bench-ai - Versions diffs - 0.1.0__tar.gz - Mend

local-bench-ai 0.1.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (334) hide show

local_bench_ai-0.1.0/PKG-INFO ADDED Viewed

@@ -0,0 +1,17 @@
+Metadata-Version: 2.4
+Name: local-bench-ai
+Version: 0.1.0
+Summary: Thin benchmark runner for OpenAI-compatible chat completion endpoints.
+Requires-Python: >=3.11
+Requires-Dist: anyio>=4
+Requires-Dist: httpx>=0.27
+Requires-Dist: langdetect>=1.0.9
+Requires-Dist: math-verify>=0.9.0
+Provides-Extra: build
+Requires-Dist: datasets>=2.20; extra == "build"
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: nltk>=3.9; extra == "dev"
+Requires-Dist: jinja2>=3.1; extra == "dev"
+Provides-Extra: hf
+Requires-Dist: transformers>=4.51; extra == "hf"

local_bench_ai-0.1.0/pyproject.toml ADDED Viewed

@@ -0,0 +1,50 @@
+[build-system]
+requires = ["setuptools>=68", "wheel"]
+build-backend = "setuptools.build_meta"
+[project]
+name = "local-bench-ai"
+version = "0.1.0"
+description = "Thin benchmark runner for OpenAI-compatible chat completion endpoints."
+requires-python = ">=3.11"
+dependencies = [
+  "anyio>=4",
+  "httpx>=0.27",
+  "langdetect>=1.0.9",
+  "math-verify>=0.9.0",
+]
+[project.scripts]
+localbench = "localbench.cli:main"
+localbench-monitor = "localbench.monitor_cli:main"
+[project.optional-dependencies]
+build = ["datasets>=2.20"]
+dev = ["pytest>=8", "nltk>=3.9", "jinja2>=3.1"]
+hf = ["transformers>=4.51"]
+[tool.setuptools.package-dir]
+"" = "src"
+[tool.setuptools.packages.find]
+where = ["src"]
+[tool.setuptools.package-data]
+localbench = [
+  "data/board_sources.json",
+  "data/licenses/*",
+  "data/suites/tiny-smoke-v1/*",
+  "data/suites/core-text-v1/*",
+  "data/suites/core-text-v1/LICENSES/*",
+  "scorers/ifbench/NOTICE",
+  "scorers/bfcl/NOTICE",
+  "scorers/bfcl_multi_turn/NOTICE",
+  "submissions/schemas/*.json",
+]
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+addopts = ["-ra", "--strict-config", "--strict-markers"]
+markers = [
+  "wsl: requires WSL2 AppWorld harness and bubblewrap",
+]

local_bench_ai-0.1.0/setup.cfg ADDED Viewed

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build =
+tag_date = 0

local_bench_ai-0.1.0/src/local_bench_ai.egg-info/PKG-INFO ADDED Viewed

@@ -0,0 +1,17 @@
+Metadata-Version: 2.4
+Name: local-bench-ai
+Version: 0.1.0
+Summary: Thin benchmark runner for OpenAI-compatible chat completion endpoints.
+Requires-Python: >=3.11
+Requires-Dist: anyio>=4
+Requires-Dist: httpx>=0.27
+Requires-Dist: langdetect>=1.0.9
+Requires-Dist: math-verify>=0.9.0
+Provides-Extra: build
+Requires-Dist: datasets>=2.20; extra == "build"
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: nltk>=3.9; extra == "dev"
+Requires-Dist: jinja2>=3.1; extra == "dev"
+Provides-Extra: hf
+Requires-Dist: transformers>=4.51; extra == "hf"

local_bench_ai-0.1.0/src/local_bench_ai.egg-info/SOURCES.txt ADDED Viewed

@@ -0,0 +1,332 @@
+pyproject.toml
+src/local_bench_ai.egg-info/PKG-INFO
+src/local_bench_ai.egg-info/SOURCES.txt
+src/local_bench_ai.egg-info/dependency_links.txt
+src/local_bench_ai.egg-info/entry_points.txt
+src/local_bench_ai.egg-info/requires.txt
+src/local_bench_ai.egg-info/top_level.txt
+src/localbench/__init__.py
+src/localbench/__main__.py
+src/localbench/_requests.py
+src/localbench/_response.py
+src/localbench/_scoring.py
+src/localbench/_suite.py
+src/localbench/_types.py
+src/localbench/budget_forcing.py
+src/localbench/campaign.py
+src/localbench/campaign_checkpoints.py
+src/localbench/campaign_records.py
+src/localbench/cli.py
+src/localbench/exit_codes.py
+src/localbench/lane_conformance.py
+src/localbench/manifest.py
+src/localbench/monitor_cli.py
+src/localbench/monitor_records.py
+src/localbench/monitoring.py
+src/localbench/orchestrate.py
+src/localbench/persistence.py
+src/localbench/prompt_rendering.py
+src/localbench/reasoning_leaks.py
+src/localbench/reasoning_registry.py
+src/localbench/release_test.py
+src/localbench/run_plan.py
+src/localbench/run_schema.py
+src/localbench/runner.py
+src/localbench/suite_bundle.py
+src/localbench/suite_errors.py
+src/localbench/suite_release.py
+src/localbench/suite_resolver.py
+src/localbench/suite_verify.py
+src/localbench/supervisor.py
+src/localbench/tc_json_v1_runner.py
+src/localbench/coding_exec/__init__.py
+src/localbench/coding_exec/extract.py
+src/localbench/coding_exec/orchestrate.py
+src/localbench/coding_exec/program.py
+src/localbench/coding_exec/runner.py
+src/localbench/coding_exec/sandbox.py
+src/localbench/coding_exec/score.py
+src/localbench/data/board_sources.json
+src/localbench/data/licenses/MIT.txt
+src/localbench/data/licenses/ODC-BY-1.0.txt
+src/localbench/data/suites/core-text-v1/ATTRIBUTION.md
+src/localbench/data/suites/core-text-v1/CHANGES.md
+src/localbench/data/suites/core-text-v1/NOTICE
+src/localbench/data/suites/core-text-v1/SCORECARD.json
+src/localbench/data/suites/core-text-v1/SHA256SUMS
+src/localbench/data/suites/core-text-v1/SOURCE_REVISIONS.md
+src/localbench/data/suites/core-text-v1/ifbench.jsonl
+src/localbench/data/suites/core-text-v1/itemsets.lock.json
+src/localbench/data/suites/core-text-v1/mmlu_pro.jsonl
+src/localbench/data/suites/core-text-v1/suite.json
+src/localbench/data/suites/core-text-v1/tc_json_v1.jsonl
+src/localbench/data/suites/core-text-v1/LICENSES/BFCL-Apache-2.0
+src/localbench/data/suites/core-text-v1/LICENSES/IFBench-ODC-BY-1.0
+src/localbench/data/suites/core-text-v1/LICENSES/IFEval-Apache-2.0
+src/localbench/data/suites/core-text-v1/LICENSES/MMLU-Pro-MIT
+src/localbench/data/suites/tiny-smoke-v1/ifbench.jsonl
+src/localbench/data/suites/tiny-smoke-v1/itemsets.lock.json
+src/localbench/data/suites/tiny-smoke-v1/mmlu_pro.jsonl
+src/localbench/data/suites/tiny-smoke-v1/suite.json
+src/localbench/kld/__init__.py
+src/localbench/kld/churn.py
+src/localbench/kld/parse.py
+src/localbench/kld/run.py
+src/localbench/probe/__init__.py
+src/localbench/probe/__main__.py
+src/localbench/probe/_point_biserial.py
+src/localbench/probe/discrimination.py
+src/localbench/probe/gates.py
+src/localbench/providers/__init__.py
+src/localbench/providers/_anthropic.py
+src/localbench/providers/_base.py
+src/localbench/providers/_openai.py
+src/localbench/scorers/__init__.py
+src/localbench/scorers/_reasoning.py
+src/localbench/scorers/lcb.py
+src/localbench/scorers/math_numeric.py
+src/localbench/scorers/math_symbolic.py
+src/localbench/scorers/mcq.py
+src/localbench/scorers/ruler.py
+src/localbench/scorers/bfcl/NOTICE
+src/localbench/scorers/bfcl/__init__.py
+src/localbench/scorers/bfcl/_checker.py
+src/localbench/scorers/bfcl/_checker_values.py
+src/localbench/scorers/bfcl/_parser.py
+src/localbench/scorers/bfcl/_prompt.py
+src/localbench/scorers/bfcl/_types.py
+src/localbench/scorers/bfcl/scorer.py
+src/localbench/scorers/bfcl_multi_turn/NOTICE
+src/localbench/scorers/bfcl_multi_turn/__init__.py
+src/localbench/scorers/bfcl_multi_turn/_backend.py
+src/localbench/scorers/bfcl_multi_turn/_executor.py
+src/localbench/scorers/bfcl_multi_turn/_parser.py
+src/localbench/scorers/bfcl_multi_turn/_prompt.py
+src/localbench/scorers/bfcl_multi_turn/_sandbox.py
+src/localbench/scorers/bfcl_multi_turn/_types.py
+src/localbench/scorers/bfcl_multi_turn/scorer.py
+src/localbench/scorers/ifbench/NOTICE
+src/localbench/scorers/ifbench/__init__.py
+src/localbench/scorers/ifbench/_checks_count.py
+src/localbench/scorers/ifbench/_checks_custom.py
+src/localbench/scorers/ifbench/_checks_format.py
+src/localbench/scorers/ifbench/_checks_ratio.py
+src/localbench/scorers/ifbench/_checks_repeat.py
+src/localbench/scorers/ifbench/_checks_sentence.py
+src/localbench/scorers/ifbench/_checks_words.py
+src/localbench/scorers/ifbench/_types.py
+src/localbench/scorers/ifbench/_util.py
+src/localbench/scorers/ifbench/instructions.py
+src/localbench/scorers/ifbench/scorer.py
+src/localbench/scorers/ifeval/__init__.py
+src/localbench/scorers/ifeval/_checks_format.py
+src/localbench/scorers/ifeval/_checks_keywords.py
+src/localbench/scorers/ifeval/_checks_length.py
+src/localbench/scorers/ifeval/_checks_misc.py
+src/localbench/scorers/ifeval/_shared.py
+src/localbench/scorers/ifeval/_types.py
+src/localbench/scorers/ifeval/_util.py
+src/localbench/scorers/ifeval/instructions.py
+src/localbench/scorers/ifeval/scorer.py
+src/localbench/scorers/tc_json_v1/__init__.py
+src/localbench/scorers/tc_json_v1/_parser.py
+src/localbench/scorers/tc_json_v1/_types.py
+src/localbench/scorers/tc_json_v1/scorer.py
+src/localbench/scoring/__init__.py
+src/localbench/scoring/axes.py
+src/localbench/scoring/axis_status.py
+src/localbench/scoring/benchmark_registry.py
+src/localbench/scoring/board.py
+src/localbench/scoring/board_manifest.py
+src/localbench/scoring/board_scoring.py
+src/localbench/scoring/board_sources.py
+src/localbench/scoring/board_support.py
+src/localbench/scoring/board_systems.py
+src/localbench/scoring/board_types.py
+src/localbench/scoring/bootstrap.py
+src/localbench/scoring/metadata.py
+src/localbench/scoring/paired_delta.py
+src/localbench/scoring/public_rescore.py
+src/localbench/scoring/scorecard.py
+src/localbench/scoring/signed_score.py
+src/localbench/scoring/subgroups.py
+src/localbench/scoring/tc_json_conformance.py
+src/localbench/scoring/web.py
+src/localbench/scoring/agentic_exec/__init__.py
+src/localbench/scoring/agentic_exec/adapter.py
+src/localbench/scoring/agentic_exec/benchmark.py
+src/localbench/scoring/agentic_exec/block_introspect.py
+src/localbench/scoring/agentic_exec/block_parser.py
+src/localbench/scoring/agentic_exec/chat_client.py
+src/localbench/scoring/agentic_exec/config.py
+src/localbench/scoring/agentic_exec/env_host.py
+src/localbench/scoring/agentic_exec/funnel.py
+src/localbench/scoring/agentic_exec/hashing.py
+src/localbench/scoring/agentic_exec/loop_config.py
+src/localbench/scoring/agentic_exec/loop_types.py
+src/localbench/scoring/agentic_exec/model_client.py
+src/localbench/scoring/agentic_exec/observations.py
+src/localbench/scoring/agentic_exec/parser.py
+src/localbench/scoring/agentic_exec/prompt.py
+src/localbench/scoring/agentic_exec/protocol.py
+src/localbench/scoring/agentic_exec/protocol_c_loop.py
+src/localbench/scoring/agentic_exec/runner.py
+src/localbench/scoring/agentic_exec/runner_bootstrap.py
+src/localbench/scoring/agentic_exec/sandbox.py
+src/localbench/scoring/agentic_exec/sandbox_protocol.py
+src/localbench/scoring/agentic_exec/score.py
+src/localbench/scoring/agentic_exec/scripted_agent.py
+src/localbench/scoring/agentic_exec/stub_appworld.py
+src/localbench/scoring/agentic_exec/task_pool.py
+src/localbench/scoring/agentic_exec/types.py
+src/localbench/scoring/agentic_exec/wsl_bridge.py
+src/localbench/scoring/agentic_exec/wsl_worker.py
+src/localbench/serving/__init__.py
+src/localbench/serving/assembly.py
+src/localbench/serving/bench.py
+src/localbench/serving/fingerprint.py
+src/localbench/serving/job_object.py
+src/localbench/serving/llama_cpp.py
+src/localbench/serving/model_artifact.py
+src/localbench/serving/options.py
+src/localbench/serving/process.py
+src/localbench/serving/provenance.py
+src/localbench/serving/readiness.py
+src/localbench/serving/runner.py
+src/localbench/serving/teardown.py
+src/localbench/submissions/__init__.py
+src/localbench/submissions/archive.py
+src/localbench/submissions/attestation.py
+src/localbench/submissions/bundle.py
+src/localbench/submissions/bundle_input.py
+src/localbench/submissions/canon.py
+src/localbench/submissions/client.py
+src/localbench/submissions/contracts.py
+src/localbench/submissions/crypto.py
+src/localbench/submissions/dedup.py
+src/localbench/submissions/divergence.py
+src/localbench/submissions/foundation.py
+src/localbench/submissions/foundation_scores.py
+src/localbench/submissions/keys.py
+src/localbench/submissions/origin.py
+src/localbench/submissions/ports.py
+src/localbench/submissions/projection.py
+src/localbench/submissions/provenance.py
+src/localbench/submissions/rescore.py
+src/localbench/submissions/status_update.py
+src/localbench/submissions/submit_run.py
+src/localbench/submissions/submit_run_inputs.py
+src/localbench/submissions/submit_run_output.py
+src/localbench/submissions/trust.py
+src/localbench/submissions/validate.py
+src/localbench/submissions/verify.py
+src/localbench/submissions/schemas/__init__.py
+src/localbench/submissions/schemas/accepted_result_projection_v1.schema.json
+src/localbench/submissions/schemas/result_bundle_v1.schema.json
+src/localbench/submissions/schemas/submission_envelope_v1.schema.json
+src/localbench/submissions/schemas/submission_item_v1.schema.json
+src/localbench/submissions/schemas/submission_manifest_v1.schema.json
+src/localbench/submissions/schemas/submission_verification_v1.schema.json
+src/localbench/submissions/schemas/suite_release_manifest_v1.schema.json
+tests/test_agentic_appworld_adapter.py
+tests/test_agentic_failure_policies.py
+tests/test_agentic_hash_stability.py
+tests/test_agentic_parser.py
+tests/test_agentic_protocol_schema.py
+tests/test_agentic_score_asr.py
+tests/test_agentic_scripted_runner.py
+tests/test_agentic_task_pool.py
+tests/test_agentic_wsl_bridge.py
+tests/test_agentic_wsl_bridge_acceptance.py
+tests/test_appworld_c_funnel_units.py
+tests/test_appworld_protocol_c_acceptance.py
+tests/test_appworld_protocol_c_gauntlet.py
+tests/test_appworld_protocol_c_units.py
+tests/test_appworld_sandbox_acceptance.py
+tests/test_appworld_sandbox_units.py
+tests/test_axes_registry.py
+tests/test_axis_measurement_status.py
+tests/test_bfcl.py
+tests/test_bfcl_multi_turn.py
+tests/test_board.py
+tests/test_board_cli.py
+tests/test_board_manifest.py
+tests/test_board_provenance.py
+tests/test_budget_forcing.py
+tests/test_build_v1_bfcl_multi_turn.py
+tests/test_build_v1_mmlu_pro.py
+tests/test_campaign.py
+tests/test_campaign_contracts.py
+tests/test_cheat_proxy.py
+tests/test_cli_axis_measurement_status.py
+tests/test_cli_bench_exit_codes.py
+tests/test_coding_exec_harness.py
+tests/test_coding_exec_orchestrate.py
+tests/test_coding_exec_sandbox.py
+tests/test_distribution_cli.py
+tests/test_gemma_reasoning_mode.py
+tests/test_genmath_gen.py
+tests/test_genmath_private.py
+tests/test_ifbench.py
+tests/test_ifeval.py
+tests/test_kld.py
+tests/test_lane_conformance.py
+tests/test_lane_enforcement.py
+tests/test_lcb.py
+tests/test_math_genmath_parity.py
+tests/test_math_numeric.py
+tests/test_math_symbolic.py
+tests/test_math_symbolic_robustness.py
+tests/test_mcq.py
+tests/test_monitoring.py
+tests/test_online_distribution.py
+tests/test_orchestrate.py
+tests/test_orchestrate_agentic.py
+tests/test_probe_discrimination.py
+tests/test_probe_gates.py
+tests/test_provider_orchestrate.py
+tests/test_provider_profiles.py
+tests/test_reasoning_registry.py
+tests/test_release_test.py
+tests/test_response.py
+tests/test_response_wrapper_scoring.py
+tests/test_run_plan.py
+tests/test_run_record_distribution_schema.py
+tests/test_runner.py
+tests/test_runner_reasoning.py
+tests/test_scorecard.py
+tests/test_scoring_aggregate.py
+tests/test_scoring_reasoning_strip.py
+tests/test_scoring_v1.py
+tests/test_serving_bench.py
+tests/test_serving_provenance.py
+tests/test_serving_teardown.py
+tests/test_site_parity.py
+tests/test_submission_slice_d1_migration.py
+tests/test_suite_bundle.py
+tests/test_suite_release_manifest.py
+tests/test_suite_resolver.py
+tests/test_supervisor.py
+tests/test_tc_json_conformance_gate.py
+tests/test_tc_json_v1_items.py
+tests/test_tc_json_v1_runner.py
+tests/test_tc_json_v1_scorer.py
+tests/test_v1_bfcl_axis.py
+tests/test_v1_bfcl_items.py
+tests/test_v1_bfcl_multi_turn_axis.py
+tests/test_v1_bfcl_multi_turn_items.py
+tests/test_v1_bigcodebench_items.py
+tests/test_v1_ifbench_axis.py
+tests/test_v1_ifbench_items.py
+tests/test_v1_lcb_axis.py
+tests/test_v1_lcb_items.py
+tests/test_v1_math_axis.py
+tests/test_v1_math_items.py
+tests/test_v1_mmlu_pro_axis.py
+tests/test_v1_mmlu_pro_items.py
+tests/test_v1_ruler_axis.py
+tests/test_verdict_integrity_direct_finalize.py
+tests/test_wave3_attestation_run_id.py
+tests/test_wave3_cli_ux.py
+tests/test_web_build_data.py
+tests/test_web_scorecard.py

local_bench_ai-0.1.0/src/local_bench_ai.egg-info/dependency_links.txt ADDED Viewed

	@@ -0,0 +1 @@
1	+

local_bench_ai-0.1.0/src/local_bench_ai.egg-info/entry_points.txt ADDED Viewed

@@ -0,0 +1,3 @@
+[console_scripts]
+localbench = localbench.cli:main
+localbench-monitor = localbench.monitor_cli:main

local_bench_ai-0.1.0/src/local_bench_ai.egg-info/requires.txt ADDED Viewed

@@ -0,0 +1,15 @@
+anyio>=4
+httpx>=0.27
+langdetect>=1.0.9
+math-verify>=0.9.0
+[build]
+datasets>=2.20
+[dev]
+pytest>=8
+nltk>=3.9
+jinja2>=3.1
+[hf]
+transformers>=4.51

local_bench_ai-0.1.0/src/local_bench_ai.egg-info/top_level.txt ADDED Viewed

	@@ -0,0 +1 @@
1	+ localbench

local_bench_ai-0.1.0/src/localbench/__init__.py ADDED Viewed

	@@ -0,0 +1 @@
1	+ """Local benchmark runner package."""

local_bench_ai-0.1.0/src/localbench/__main__.py ADDED Viewed

@@ -0,0 +1,11 @@
+"""Module entry point for python -m localbench."""
+from __future__ import annotations
+import sys
+from localbench.cli import main
+if __name__ == "__main__":
+    sys.exit(main())