Synth Compass (ESQSetEncoder)

A 137,897-parameter CPU neural search compass for deterministic program synthesis.

Synth combines execution-signature quotienting with branch-preserving symbolic search. The ESQSetEncoder checkpoint predicts operator, lambda, and program-length priors to reorder search visitation. It does not predict program outputs and does not prune valid symbolic branches.

  • Author: Kshitij Pal
  • Organization: TENS LAB β€” https://tenslab.dev
  • Source: https://github.com/Kshitijpalsinghtomar/synth
  • License: Apache-2.0 (first-party source code and model weights)
  • Checkpoint: esq_v3_converged.pt
  • Parameters: 137,897 total (78,664 used by trained prior heads; 59,233 in a legacy candidate-ranker branch not optimized by this objective)
  • Inference: CPU; no GPU is required

Repository files

File Purpose
esq_v3_converged.pt Packaged PyTorch checkpoint
model_card.json Architecture, training, provenance and limitations metadata
esq_v3_converged.pt.sha256 Expected SHA-256 digest for the checkpoint
LICENSE Apache License 2.0

Checkpoint integrity

Expected SHA-256:

a0f03d5066d86022d932699c1a6dfe68957d56131a38e7816e2085bda2000d41

Verify after downloading:

sha256sum -c esq_v3_converged.pt.sha256

The expected digest identifies the exact packaged bytes. It is an integrity check, not a malware scan.

Download and use with Synth

The Hub stores the checkpoint, model card, and SHA-256 sidecar at the repository root. Download all three together into Synth's checkpoints/ directory so the package's loader can verify them as one artifact:

hf download Kshitijpalsinghtomar/synth-compass esq_v3_converged.pt model_card.json esq_v3_converged.pt.sha256 --local-dir checkpoints

From a clean Synth checkout, install the neural extras and run the guided search mode:

pip install -e ".[neural]"
synth sample_task.json --mode guided --checkpoint checkpoints/esq_v3_converged.pt

Use the Synth NeuralGuide loader rather than calling torch.load on arbitrary files yourself. It verifies the checkpoint SHA-256 against both the model card and sidecar, then uses restricted weights_only=True deserialization.

PyTorch security: CVE-2026-24747 affects the weights_only unpickler in PyTorch versions earlier than 2.10.0. Synth's loader permits only the exact hash-pinned bundled checkpoint on those older runtimes; custom or third-party checkpoints require PyTorch 2.10.0 or newer. Never disable weights_only to make an untrusted checkpoint load. SHA-256 confirms byte identity with the expected digest; it is not a malware scan. Hugging Face also runs file and pickle-import scans, but those checks are best-effort and are not a safety guarantee. See the Hugging Face pickle-scanning guidance and PyTorch security advisory.

Architecture

ESQSetEncoder consumes a fixed-shape task representation with input feature shape [3, 4, 37], an 8-dimensional variable-position embedding, value/example encoders, and a task projection. Three trained heads predict:

  • Operator priors: 15 DeepCoder operators.
  • Lambda priors: 19 supported lambda expressions.
  • Program-length prior: lengths 1–6.

The training objective was binary cross-entropy for operator and lambda heads plus weighted cross-entropy for length. Inference uses PyTorch and runs on CPU.

The learned scores affect only enumeration order. Search remains symbolic, bounded, and branch-preserving. The model is not a program-output oracle.

Locked V6 benchmark: strict depth-4 zero-overlap cohort

The V6 structural stress test used 196 eligible tasks and 8,869 fresh cases. No AST structural family in the evaluation cohort overlapped with training.

Configuration Complete task matches Full-task accuracy Fresh-case accuracy
V1 symbolic 60/196 30.61% 70.03%
V3 guided 62/196 31.63% 70.77%
V5 ablation 70/196 35.71% 74.36%
V5 full 71/196 36.22% 74.34%

The V5 full arm is 5.61 percentage points above V1 symbolic on complete task matches. This is an observed comparison across configurations, not an isolated causal estimate of the oracle component. The full and ablation arms differ by 0.51 percentage points in complete-task accuracy.

This is a structural-family-held-out synthetic evaluation under strict depth-4 semantics with a local file-backed synthetic oracle. It is not an official ExeDec benchmark or a human-oracle evaluation.

Standard external DeepCoder splits

These evaluations are separate from the V6 structural stress test:

Split System Complete fresh-behaviour matches Fresh-case accuracy
SWITCH_CONCEPT_ORDER (ESQ-03) Synth 144/180 (80.0%) 88.9% (7,947/8,937)
LENGTH_GENERALIZATION (V3) Synth 118/180 (65.6%) 78.7%
LENGTH_GENERALIZATION Pure symbolic 108/180 (60.0%) 75.2%

The LENGTH_GENERALIZATION confirmation used the original V3 checkpoint 1e5e086f, distinct from the packaged checkpoint a0f03d50. Do not attribute that result to the packaged checkpoint.

Training and provenance notes

The canonical model_card.json records the architecture, training split, optimizer, epoch history, held-out prediction metrics, source-data digests, and limitations. The historical release rebuild has partial provenance: corpus and supervision hashes are known, but the exact training-entrypoint hash and runtime environment were not retained. Exact training reproduction is therefore not claimed.

Usage

This checkpoint is intended to be loaded by the Synth package and used by its guided search mode. See the source repository for installation, command-line examples, runtime compatibility restrictions, and the complete synthesis pipeline:

https://github.com/Kshitijpalsinghtomar/synth

Citation

@software{synth_esq_dsl_2026,
  author       = {Kshitij Pal},
  title        = {Synth: Execution-Signature Quotienting for Deterministic Program Synthesis},
  year         = {2026},
  organization = {TENS LAB},
  url          = {https://tenslab.dev},
  version      = {1.0.0},
  license      = {Apache-2.0}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support