HuCore-68M Symbolic Topology Controller

One researcher. One RTX 4070. One custom architecture. 1.0000.

HuCore-68M is a compact symbolic topology controller trained to turn elementary word problems into structured slot/skeleton proposals.

It is not a chatbot. It is a small control model for one job:

word problem -> numeric slots -> symbolic skeleton -> checked evaluator

Example:

you have 7 apples then get 14 more and give 2 away
-> a 7 b 14 c 2 : a + b - c

This public release is weights-only and includes the evaluation test sets used in the report.

TL;DR

  • 68.43M parameters
  • Custom HuCore architecture: not LLaMA, not Qwen, not Gemma, not a public-model fork
  • Controlled topology holdout: 240 / 240
  • Local training hardware: one NVIDIA RTX 4070 Laptop GPU
  • Release format: weights + test sets + reports, no implementation code
  • License: CC-BY-NC-4.0

What It Does

HuCore-68M extracts a compact symbolic form from simple arithmetic word problems:

Input pattern Target output
add / subtract word problem a b c : a + b - c
equal groups a b : a x b
equal sharing a b : a / b
comparison wording a b : a + b

The model proposes the structure. A deterministic evaluator computes the final value.

Scoreboard

Controlled topology holdout:

Dataset Total Slot Skeleton Value
controlled_topology_holdout_240 240 1.0000 1.0000 1.0000

Stress v1:

Variant Value Exact
punctuation_flip 1.0000
whitespace_noise 1.0000
polite_prefix 0.8958
neutral_suffix 0.0708
case_noise 0.1583
distractor_suffix_number 0.0042

The stress set is included because the point is not to hide weak spots. The point is to make the boundary visible.

Repository Contents

File Description
pytorch_model.bin Weights-only checkpoint
metrics.json Compact metric summary
testsets/controlled_topology_holdout_240.jsonl Controlled evaluation set
testsets/stress_v1_1440.jsonl Stress evaluation set
reports/benchmark_pack_v1_controlled_and_stress.md Evaluation report
TRAINING_ENVIRONMENT.md Local training environment
MODEL_CARD.zh-CN.md Chinese model card
LICENSE.md CC-BY-NC-4.0 license note

This artifact does not include source code, tokenizer files, architecture config, training scripts, or training data.

Why 68M?

This release tests a simple claim:

For controlled symbolic topology extraction, a small custom controller can be enough.

More parameters are not automatically better if the task is a narrow control problem. HuCore-68M is deliberately small so the extraction behavior can be inspected, stress-tested, and improved without hiding everything inside a large general model.

Intended Use

Use this artifact for research on:

  • compact symbolic controllers;
  • slot/skeleton extraction;
  • boundary-first model behavior;
  • deterministic evaluator pipelines;
  • small-model stress testing.

Not For

HuCore-68M is not intended as:

  • a general chat assistant;
  • a full GSM8K solver;
  • a bilingual math model;
  • a commercial guardrail;
  • a drop-in Transformers model.

License

All files in this public artifact are released under CC-BY-NC-4.0.

Commercial use, hosted service integration, proprietary product integration, or model distillation for commercial deployment requires separate permission from the author.

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. 🙋 Ask for provider support