prereasoner/prereasoner-weights

Prereasoner Runtime Weights

This repository contains the public, manifest-pinned runtime weights for Prereasoner, an interpretable table-question-answering system. Learned components produce semantic evidence: named coordinates for columns and question words, Schema.org property and class scores, and the own-data readers' readings of a question (which tables, columns, values, groups, orders and calculations it asks for). A deterministic typed AST builder makes every query from those readings and owns joins, calculation verification, rendering, and execution. No model in this bundle writes SQL, and the bundle is not a standalone text-to-SQL model.

Install

No Hugging Face account or token is required:

git clone https://github.com/ManiDoraisamy/prereasoner-data.git
cd prereasoner-data
python -m engine.fetch_weights

engine/data/weights_manifest.json pins this repository at an immutable commit and records the SHA-256 digest of every required file. The fetcher downloads into a temporary directory, verifies the complete bundle, and only then installs it.

Bundle Contents

File	Purpose
encoder.pt	State dictionary for the relational semantic readout
encoder_meta.pt	Readout allocation and constructor configuration
qwen_lora/adapter_model.safetensors	LoRA adapter for Qwen/Qwen2.5-0.5B
qwen_lora/adapter_config.json	PEFT adapter configuration
anchor_assignment.npz	Calibrated named-dimension thresholds
primitives.npz	Learned primitive-composition head
schema_property_head.pt	Calibrated Schema.org named-property evidence head
readers/stage1/	Own-data schema-item model (XLM-R base): scores each table's and column's use
readers/stage2_a/ to readers/stage2_d/	Four own-data reading models (XLM-R large); each reader.json says whether it reads the question's values beside their columns
readers/rescorer/	Re-scorer (XLM-R base): scores the readings' queries that ran
readers/choice.json	The readers' choice: a logistic model over the re-scorer's logit and how many readers build a query, fitted on held-out calibration databases; it ranks the queries and refuses below its threshold

The source repository contains the small ontology, calibration, taxonomy, and manifest artifacts. The base Qwen model is downloaded separately from its publisher. The readers' directories hold their complete fine-tuned weights (encoders in bfloat16, heads in float32) and tokenizer files. python -m engine.fetch_weights downloads only the files the source manifest lists. Prereasoner loads no SQL-generating model; the retired SQL proposer adapter (sql_proposer/) was removed from this repository on 2026-10-02 and remains in its history.

Model Boundary

The encoder adapter emits named Schema.org property probabilities, calibrated class proposals, and embeddings for column typing and the structural intent of composition and world questions. A released class may propose a coarse resolver family, but no score can authorize a table, join, calculation, or answer. Own-data questions are read by separate XLM-R readers (engine/data/readers/); a deterministic builder makes each of their readings a typed query, and no model writes SQL. The optional Gemini model serves only the chat orchestrator, starter questions and reference generation, and is not part of this bundle. Deterministic code applies ontology mapping, exact source grounding, typed constraints, abstention rules, and execution checks.

Schema.org 30.0 supplies the named semantic coordinate system. Wikidata and publisher datasets supply observations mapped into those coordinates; mutable source facts are not intended to be memorized as answers.

Evaluation And Provenance

See the source repository's model card, data card, and Spider results for component boundaries, denominators, and limitations.

The bytes in this bundle are immutable and hash-verified. The promoted Schema.org head has a machine-readable corpus, split, seed, dependency, encoder, and held-out metric manifest. The shared encoder's historical training run has less completely recorded source-corpus provenance; do not combine metrics from the two tracks. All 926 Schema.org classes are representable, but only the released calibrated subset is servable. Unsupported and under-calibrated coordinates abstain.

Intended Use

• Semantic evidence for the matching Prereasoner source revision.
• Research on deterministic SQL planning informed by named learned dimensions.
• Local table and spreadsheet question answering with inspectable plans.

Out Of Scope

• Standalone SQL or answer generation.
• Authoritative identity or entity classification.
• High-stakes medical, legal, tax, financial, or safety decisions without source-specific review.
• Using the files with an unverified or incompatible source revision.

License

The Prereasoner weight bundle is released under Apache-2.0. The base model and source datasets retain their own licenses and terms. The readers in readers/ are fine-tuned from XLM-RoBERTa (MIT) on Spider's train and dev questions (Yu et al., EMNLP 2018; CC BY-SA 4.0), BIRD train (Li et al., NeurIPS 2023; CC BY-SA 4.0) and Spider-Syn (Gan et al., ACL 2021; MIT), and on questions Gemini on Vertex AI wrote over those training databases; keep those attributions when redistributing them. The Spider test set appears in no training step. The retired sql_proposer/ adapter in earlier revisions was trained on Spider TRAIN (Yu et al., EMNLP 2018; CC BY-SA 4.0); keep that attribution when redistributing it. See THIRD_PARTY.md.