russell@unturf.com
87d9db15c7
Speedup (§3 plan): ShadowNLI._nli_batch batches forwards (ARBORIST_NLI_BATCH=64); device auto-detect (ARBORIST_NLI_DEVICE, else cuda-if-available); auto-prefer an ONNX export — bench/scripts/export_nli_onnx.py / make export-nli-onnx exports + int8-dynamic-quantizes the pinned checkpoint into ~/.arborist/models/nli/<ver>/onnx/ (operator state, NOT committed), _ensure_loaded loads model_quantized.onnx via optimum.onnxruntime (backend onnx-int8), falls back to torch silently. torch-cpu-batch1 ~120ms/pair → onnx-int8-cpu-batched ~32ms/pair (~4x); seconds on a 4090. optimum[onnxruntime] added to the [nli] extra; 24 tests. Gate-item-4 verdict at proper n: ARBORIST_NLI_SHADOW=1 make bench-qa BENCH_QA_N=1 → 223 cells (89 STRICT / 90 HYBRID / 44 UNGROUNDED; also surfaced + fixed a lone-surrogate bug). Shadow sweep over those: NLI-as- runtime-veto on STRICT has ~26% FP at θc 0.5, ~8% at θc 0.90, ~0% only at θc 0.99 — and θc 0.99 gives up most recombination recall (hard synthetic recombinations bottom out ~0.76). FAILS the §7 #12 gate on this design. Only untried path that might pass: a Phase-3 runtime hook running NLI on the verifier's actual matched clauses (1-3), not top-6-by-overlap. Until then: runtime NLI demotion stays off; the 2 fixtures stay permanent boundary markers; θc stays 0.5. Production verifier unchanged; falsification-hard stays 10/12.
124 lines
4.1 KiB
TOML
124 lines
4.1 KiB
TOML
[build-system]
|
|
requires = ["setuptools>=68"]
|
|
build-backend = "setuptools.build_meta"
|
|
|
|
[project]
|
|
name = "arborist"
|
|
version = "0.0.1"
|
|
description = "An arborist for trees and forests of cross-linked information"
|
|
readme = "README.md"
|
|
license = { text = "AGPL-3.0-only" }
|
|
requires-python = ">=3.10"
|
|
authors = [
|
|
{ name = "Russell Ballestrini", email = "russell@unturf.com" },
|
|
{ name = "foxhop" },
|
|
{ name = "TimeHexOn" },
|
|
]
|
|
dependencies = [
|
|
"httpx>=0.27",
|
|
"zstandard>=0.22",
|
|
"cryptography>=42",
|
|
]
|
|
|
|
[project.optional-dependencies]
|
|
html = [
|
|
"selectolax>=0.3",
|
|
]
|
|
wikitext = [
|
|
"mwparserfromhell>=0.6",
|
|
]
|
|
mesh = [
|
|
# httpx is already in core deps; mesh wire only depends on stdlib +
|
|
# cryptography (also core). This extras block exists as the documented
|
|
# opt-in surface even though no extra packages are required today.
|
|
]
|
|
math = [
|
|
# Symbolic algebra/calculus π* substrate (ticket #000030). SymPy is
|
|
# ~30 MB installed; pulling it into core deps would inflate every
|
|
# fresh checkout. Tests skip via pytest.importorskip when absent.
|
|
"sympy>=1.13",
|
|
]
|
|
hessian = [
|
|
# Phi_alignment_probe (ticket #000034 Phase 1a). Lanczos top-k +
|
|
# bottom-k eigendecomposition for measuring whether v7's frozen
|
|
# linear projection W aligns with the loss Hessian's low-eigenvalue
|
|
# subspace. Numpy + scipy together ~80 MB; gated separately from
|
|
# core to keep the default install lightweight. Tests skip via
|
|
# pytest.importorskip when absent. Install with:
|
|
# pip install 'arborist[hessian]'
|
|
"numpy>=1.26",
|
|
"scipy>=1.11",
|
|
]
|
|
crawler = [
|
|
# Verbatim lift from agents.ai.unturf.com/core. Off by default — the
|
|
# default test suite never imports the crawler. Install with:
|
|
# pip install 'arborist[crawler]'
|
|
# then run `make test-crawler`.
|
|
"aiohttp>=3.8",
|
|
"beautifulsoup4>=4.11",
|
|
"lxml>=4.9",
|
|
"html5lib>=1.1",
|
|
"html2text>=2024.2.26",
|
|
"miniuri>=1.1",
|
|
"feedparser>=6.0",
|
|
"Pillow>=10.0",
|
|
"cairosvg>=2.7",
|
|
"pypdf>=4.0",
|
|
]
|
|
vec = [
|
|
# Optional sqlite-vec semantic retrieval backend (ticket #000039).
|
|
# sqlite-vec ships only the loadable SQLite extension (~1 MB);
|
|
# fastembed pulls onnxruntime + tokenizers + huggingface-hub
|
|
# (~150 MB) and downloads the bge-small-en-v1.5 ONNX model
|
|
# (~130 MB) on first use. Gated separately so a fresh checkout
|
|
# stays python3.12 + venv + sqlite3. CLI surfaces `arborist embed`
|
|
# / `--backend vec` only when `sqlite_vec` imports. Install with:
|
|
# pip install 'arborist[vec]'
|
|
# (sentence-transformers is the heavier "official" embedder path
|
|
# the ticket §5 names; fastembed is the lightweight ONNX one.)
|
|
"sqlite-vec>=0.1.9",
|
|
"fastembed>=0.4",
|
|
]
|
|
nli = [
|
|
# Sentence-pair NLI for the #000049 Phase-2 *shadow* path
|
|
# (arborist/qa/nli/) — measures whether a clause-level contradiction
|
|
# veto would demote a weakly-grounded answer; never touches
|
|
# audit_mode. transformers + a CPU torch is ~600 MB installed, so it
|
|
# is gated hard out of core / dev — a fresh checkout stays
|
|
# python3.12 + venv + sqlite3, and the default test suite skips the
|
|
# NLI tests via pytest.importorskip when this extra is absent.
|
|
# Install with:
|
|
# pip install 'arborist[nli]'
|
|
# `optimum[onnxruntime]` gives the ONNX-export + int8-quantize path
|
|
# (`bench/scripts/export_nli_onnx.py`, `make export-nli-onnx`):
|
|
# `onnxruntime` on a quantized cross-encoder is ~2-4x faster on CPU
|
|
# than the torch forward path; `ShadowNLI._ensure_loaded` auto-prefers
|
|
# an export if it finds one. torch is still here because `optimum`'s
|
|
# exporter uses it, and it's the fallback when no export exists; a
|
|
# Phase-3 runtime could ship an `[nli-onnx]`-only extra (onnxruntime,
|
|
# no torch) once the export is committed/distributed (cf. [vec]).
|
|
"transformers>=4.40",
|
|
"torch>=2.2",
|
|
"sentencepiece>=0.2",
|
|
"protobuf>=4.0",
|
|
"optimum[onnxruntime]>=1.20",
|
|
]
|
|
dev = [
|
|
"pytest>=8",
|
|
"pytest-asyncio>=0.23",
|
|
"pytest-xdist>=3.5",
|
|
"arborist[html]",
|
|
"arborist[wikitext]",
|
|
"arborist[mesh]",
|
|
"arborist[crawler]",
|
|
"arborist[math]",
|
|
"arborist[hessian]",
|
|
"arborist[vec]",
|
|
]
|
|
|
|
[project.scripts]
|
|
arborist = "arborist.cli:main"
|
|
|
|
[tool.setuptools.packages.find]
|
|
where = ["."]
|
|
include = ["arborist*"]
|