pkcii/eval-impact-dolly

Dolly for OPD vs OPD + IMPACT

Data for the LLaMA2-7B → TinyLLaMA-1.1B comparison of OPD and OPD + IMPACT, with fixed, disjoint BI calibration samples. This repository contains data and provenance only, with no model weights or training results. MATH-500/GSM8K are excluded from this release.

Source and attribution

The underlying instruction/response dataset is Databricks Dolly 15k, distributed under CC BY-SA 3.0. Credit: Databricks and the Dolly data contributors.

The processed Dolly splits here are the files bundled in the IMPACT code-and-data supplement, using the MiniLLM-style prompt format. DSKD and DSKDv2 use processed Dolly for their instruction-following experiments. This is a community experiment data release, not an official checkpoint or dataset release by those authors. Exact supplement and split hashes are recorded in provenance.json; the authors' Google Drive data were not downloaded for this release.

Configs and counts

Config	Splits	Purpose
opd (default)	train 11,421; validation 256; test 500	Prepared local experiment data. validation is the fixed BI seed-42 subset, not the full original dev split.
bi	seed42 256; seed43 256; seed44 256	Fixed BI subsets for sampling stability; subsets can overlap each other.
original	train 11,435; validation 1,000; test 500	Original bundled split rows, including original duplicates/overlaps.

The experiment train split excludes 14 rows whose prompt hashes overlap test prompts. Other train duplicates remain to preserve the original recipe as closely as possible. BI samples are drawn from 972 distinct eligible dev prompts after excluding both train and test prompts. Train and each BI subset do not overlap; each BI subset and test do not overlap. The original config intentionally preserves the original overlaps for provenance and should not be substituted into a leakage-controlled experiment without auditing.

Processed columns: id, prompt, response, domain, source, source_index, prompt_sha256, prompt_format. Prompts and responses are kept as in the bundled source. prompt_format=raw means no extra BOS, chat template or newline. The training recipe tokenizes first and truncates the prompt to 256 tokens; data files retain the full prompt text.

Load

from datasets import load_dataset

opd = load_dataset("pkcii/eval-impact-dolly", "opd")
bi = load_dataset("pkcii/eval-impact-dolly", "bi")
original = load_dataset("pkcii/eval-impact-dolly", "original")

The jsonl/ folder contains the exact normalized JSONL files consumed by the local trainer, including calibration.jsonl as an alias for seed 42. Parquet files are included for the dataset viewer. original/ retains the source JSONL and corresponding Parquet files. The Parquet files use the union of original columns; topic is null where that metadata field is absent in the source train/dev splits. Original JSONL bytes are unchanged.

Models and code

• Teacher base: meta-llama/Llama-2-7b-hf, followed by local Dolly teacher SFT.
• Student base: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T.
• These exact base model IDs are listed in the original DSKD Models section. This does not identify the original authors' post-SFT teacher adapter or a released DSKDv2-trained student.
• Experiment code: aiskylimit/new_nothing, chi_eval_impact.

Limitations

This release supports training and BI analysis, not a claim of reproduced DSKDv2 scores. The three BI subsets all come from Dolly dev and measure sampling sensitivity rather than cross-dataset stability. The base models share a tokenizer; this setup does not establish a cross-tokenizer result. The test split is retained for reproducibility; the current experiment does not schedule benchmark evaluation.

License and changes

Dataset content and this adapted data release are provided under CC BY-SA 3.0, with attribution to the original Dolly contributors. Changes consist of preserving the supplied prompt format, adding normalized metadata, filtering train/test prompt overlap, selecting BI calibration subsets, and converting to JSONL/Parquet. The experiment code's software license does not replace the dataset license. No endorsement by Databricks, Microsoft, the DSKD authors or the IMPACT authors is implied.