wsp-agent-data / README.md
yemara's picture
Agent validation data
4f3f79f
|
Raw History Blame Contribute Delete
2.18 kB
metadata
configs:
  - config_name: chat_math
    data_files: data/chat_math.parquet
  - config_name: prose
    data_files: data/prose.parquet
  - config_name: mmlu_val
    data_files: data/mmlu_val.parquet
  - config_name: gsm8k_dev1k
    data_files: data/gsm8k_dev1k.parquet

Data used by the WSP agent

Every piece of data the agent ran the model on while repairing the 3-bit Nemotron-3.5-Lightning-30B-A3B checkpoints in Repairing a Quantized LLM using Weight Space Programming (wsp-ptq), exactly as its lab loaded it: the same items in the same order. The agent used it to find problems and to keep or revert edits. None of it set weight values. No test data is included.

Config What Rows Used for
chat_math the NVFP4 model's greedy answers to 256 GSM8K train questions (chat template, thinking off) 256 KL to NVFP4 (first 96, in_kl_set); vLLM NLL of the answers (all 256)
prose WikiText-2 (raw) validation in 1,024-token windows 259 perplexity (all); KL to NVFP4 (first 48, in_kl_set)
mmlu_val MMLU validation + dev 1,816 0-shot letter-likelihood accuracy and KL
gsm8k_dev1k GSM8K train questions, seed-0 permutation indices 1,480–2,479 1,000 greedy generation accuracy

About 370K tokens in total. The agent also measured KL on 32 held-out code files (29K tokens), which are not included. Token ids are for the Nemotron-3.5-Lightning tokenizer. KL on chat_math is scored on the answer tokens; on prose, from token 32 on. Localizing problems used the first 32 items of chat_math and the first 16 of prose.

The layer-42 edit on the well-calibrated checkpoint also used per-channel RMS activations taken from that checkpoint's own GPTQ calibration statistics; those are not part of this dataset.

Sources and licenses

GSM8K (MIT), MMLU (MIT), WikiText-2 (CC BY-SA 3.0), answers generated by nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 (OpenMDW-1.1). build_dataset.py rebuilt this from the lab's inputs.