HimanshuPathak/glm47-synth-v1-dataset

GLM47 Synth V1 Dataset

This repository contains 260 verified Aider-format supervised fine-tuning rows. The dataset has ten synthetic variants for each of 26 C++ task families and was structured for an exact 100-epoch memorization experiment.

Dataset structure

The dataset has one configuration (default) and one split (train):

Every row contains these top-level fields:

• —label: unique row label
• —task_id: unique synthetic task identifier
• —messages: nine role/content messages in aider-chat-v1 format
• —metadata: task family, variant, editable files, verification hashes, and related technical fields

The message-role sequence is system, user, assistant, user, assistant, user, assistant, user, assistant. Tests are not model-visible. The verification audit records that every final assistant target was replayed against its corresponding C++ test suite.

Task families

The 26 source task families are:

all-your-base, allergies, bank-account, binary-search-tree, circular-buffer, clock, complex-numbers, crypto-square, diamond, dnd-character, gigasecond, grade-school, kindergarten-garden, knapsack, linked-list, meetup, parallel-letter-frequency, perfect-numbers, phone-number, queen-attack, robot-name, space-age, spiral-matrix, sublist, yacht, and zebra-puzzle.

Each family has ten variants.

Supporting metadata

• —manifest.json describes the dataset format, counts, training recipe, tokenizer constraints, and technical provenance.
• —ROW_CATALOG.json contains the ordered row catalog and integrity hashes.
• —AUDIT.json contains the verification summary and per-row verification results.

The primary data file has SHA-256:

3472d76169e52bd0859c181d63de24a060c4c7f2d3d8a004ceb6090498f1ddc1

Intended use

The dataset is intended for supervised fine-tuning and controlled memorization experiments with Aider-style software-editing conversations. The recorded recipe uses 100 epochs, a global batch size of 20, and 1,300 optimizer steps. The manifest records tokenizer checks against zai-org/GLM-4.7-Flash with a 4,096-token sequence length; all 260 serialized examples were within that limit.

Limitations

• —This is a small synthetic dataset with a single training split; it is not a general-purpose evaluation benchmark.
• —It intentionally targets exact memorization and should not be treated as evidence of broad coding generalization.
• —The task-suite sources, tests, rebuild scripts, and independent verifier are not included in this dataset repository; their hashes and verification results are retained in the supporting metadata.
• —No explicit license was discovered in the inspected source repository. No license is asserted here.