Datasetpublic

TokenBender/glm47-synth-v1-dataset

Synth v1 Dataset This package contains 260 verified Aider-format SFT rows: ten synthetic variants for each of the 26 source task families. It is intentionally built for an exact 100-epoch memorization experiment. The training file is sft/train.jsonl. Every row uses the same nine-message aider-chat-v1 structure as the successful SFT-v5 package. Tests are not model-visible; the independent verifier replays each final assistant target against its source C++ test suite.… See the full description on the dataset page: https://huggingface.co/datasets/TokenBender/glm47-synth-v1-dataset.

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes10downloads

Dataset Card

Synth v1 Dataset

This package contains 260 verified Aider-format SFT rows: ten synthetic variants for each of the 26 source task families. It is intentionally built for an exact 100-epoch memorization experiment.

The training file is sft/train.jsonl. Every row uses the same nine-message aider-chat-v1 structure as the successful SFT-v5 package. Tests are not model-visible; the independent verifier replays each final assistant target against its source C++ test suite.

Rebuild

bash

python3 scripts/build_synth_memorization_dataset.py \
  --tokenizer-json /path/to/GLM-4.7-Flash/tokenizer.json

Verify

bash

python3 scripts/verify_synth_memorization_dataset.py

Train with the existing runner

bash

MILES_CPP_AUTO_PREPARE_DATA=0 \
MILES_CPP_DATA_DIR=/path/to/synth-v1-dataset \
MILES_SFT_NUM_EPOCH=100 \
MILES_GLOBAL_BATCH_SIZE=20 \
MILES_ROLLOUT_BATCH_SIZE=20 \
bash examples/sft.sh