KYS-Configs

Everything needed to re-run Know Your Sources: Data Selection Matters when Rewriting for Data-Constrained Pretraining end to end: the rewriting prompts, the vLLM generation settings, the Nanotron training configs, the shared initial weights, and the full evaluation harness with all raw results.

Layout

prompts/     the rewriting, annotation and judging prompts
vllm/        the rewriting workers and their SLURM launchers
nanotron/    the 18 training configs + shared init weights
eval/        LightEval tasks, launchers, converter, and 276 raw result JSONs

prompts/

Production rewriting prompts — the six templates that built the released corpora:

File	Used by
wikipedia_style_rephrasing_grounded.md	QUALITY-FIRST, DIVERSITY-ORIENTED, DISAGREEMENT-AWARE, REWIRE-INSPIRED
distill_prompt.txt	the same four, as the second pass
wrap_prompts.json	WRAP-INSPIRED — four keys: easy, hard, wiki, qa

The wikipedia-style template is the grounded variant: the upstream text plus an explicit instruction not to add information that is not stated in or inferable from the source. Each WRAP style ends with "\n\nPassage:\n" and the document is concatenated, not substituted. Style assignment is deterministic: np.random.default_rng([42, shard_index]).integers(0, 4, n_rows) over ["easy", "hard", "wiki", "qa"], recorded per row in a wrap_style column.

Supporting prompts:

• haiku_annotation_prompt.txt — the five-criterion rubric Claude Haiku 4.5 applied to the 50,427 documents behind the quality scorer (Appendix A).
• hallucination_judge_prompt.txt — the factual-consistency judge used for the prompt comparison (Appendix C).

prompts/finephrase/ — the 14 pilot templates, reconstructed and verified. See the README in that directory; they carry caveats the production prompts do not.

vllm/

Rewriting used the offline vllm.LLM Python API, not a server — 8 independent single-GPU processes data-parallel over shards. There is no endpoint or port to configure.

LLM(model=..., tensor_parallel_size=1, dtype="bfloat16",
    gpu_memory_utilization=<0.85 | 0.90>, max_model_len=32768)

max_new = min(4096, 32768 - n_input_tokens)
SamplingParams(temperature=0, top_p=1.0, max_tokens=max(1, max_new))

Generator	Qwen/Qwen2.5-7B-Instruct, revision a09a35458c702b33eeacc393d103063234e8bc28
vLLM	0.22.0
torch / transformers	2.11.0+cu130 / 5.9.0
--input-drop	30720 — longer templated inputs are skipped and marked status = 0
Hardware	1× H100 per worker, array of 8

⚠️ gpu_memory_utilization is inconsistent in the sources. README_07_rewrite.md documents 0.90; sbatch_template_rewrite.sh passes 0.85. The logs show both — 0.90 for the June 2026 runs (quality-first, diversity-first, wrap, lambda05) and 0.85 for the July 2026 runs. Under greedy decoding this does not affect outputs, but the README value is wrong for half the runs.

⚠️ No sampling seed was set. SamplingParams carries no seed, and the engine logs seed=0. Reproducibility rests on temperature=0 alone; bitwise reproducibility is not claimed.

nanotron/

The 18 configs that actually produced the released checkpoints — one authored YAML per (setting, seed), not templated:

config_{quality-base, quality-first, diversity-first, wrap, rewrite,
        signal-disagreement-lambda05}_10B_1.5B{, _seed43, _seed44}.yaml

The unsuffixed file is seed 42. Settings differ only by data_stages[0].data.dataset.dataset_folder; seeds differ by general.seed, general.run, checkpoints_path and resume_checkpoint_path. data_stages[*].data.seed is 42 throughout.

⚠️ The Nanotron working tree was dirty at training time — the commit that was checked out does not contain these configs. Provenance by commit hash is not possible, which is why they ship as files. A separate nanotron-kys reproduction pack exists in the project but produced no released checkpoint; it is deliberately not included here.

launch_nvl_generic.sbatch and run_nanotron.sh are the SLURM launcher and the torch.distributed.run --nproc_per_node 4 wrapper.

nanotron/init/ — the shared initial weights

seed42/, seed43/, seed44/: the step-0 checkpoints, converted to HF format, weights only.

This is what makes the comparison controlled. All six settings at a given seed load the same initial parameters, and since there is no dropout and the data order is fixed, the seed controls initialization alone. Row-parallel projections were scaled by 1/√(2L) with L = 28; all other weights drawn from N(0, 0.02).

from transformers import AutoModelForCausalLM
from huggingface_hub import snapshot_download
p = snapshot_download("blab-jhu/KYS-Configs", repo_type="dataset",
                      allow_patterns=["nanotron/init/seed42/*"])
m = AutoModelForCausalLM.from_pretrained(f"{p}/nanotron/init/seed42")

eval/

File	What it is
rewrite_tasks.py	the custom LightEval task file — rw_hellaswag, rw_piqa, rw_arc:easy, rw_arc:challenge, rw_openbookqa, rw_commonsense_qa, rw_siqa, rw_winogrande, rw_boolq, rw_sciq, rw_lambada, rw_mmlu:<57 subsets>
env.sh	task strings (TASKS_CS, TASKS_MMLU), TOKENS_PER_STEP=2097152
run_eval.slurm, run_eval_mmlu.slurm, run_arcc.slurm, run_extra.slurm	the launchers
ckpt_list*.txt	the (setting, step) work lists the arrays index into
convert_to_hf.py	Nanotron(tp=1, pp=1) → LlamaForCausalLM converter
results/downstream/	276 raw LightEval result JSONs
appendix_f_per_task.csv	the 361-row table backing Appendix F

acc_norm is LoglikelihoodAcc(logprob_normalization=LogProbTokenNorm()) — continuation-token length normalization, so longer options are not penalized.

The converter is verified

convert_to_hf.py produced both the paper's evaluation checkpoints and the released model repos. Re-running LightEval on a freshly converted checkpoint reproduces the stored numbers exactly:

Task	Metric	Stored	Re-run	Delta
rw_piqa|0	acc_norm	0.710555	0.710555	0.000e+00
rw_hellaswag|0	acc_norm	0.486158	0.486158	0.000e+00

⚠️ The Llama-2 tokenizer used here requires tokenizers >= 0.20; older versions fail to load it with data did not match any variant of untagged enum ModelWrapper.

The rest of the release

Repo	What it holds
KYS-1.5B-Quality-Base	QUALITY-BASE — non-rewritten baseline
KYS-1.5B-Quality-First	QUALITY-FIRST
KYS-1.5B-Diversity-Oriented	DIVERSITY-ORIENTED
KYS-1.5B-Disagreement-Aware	DISAGREEMENT-AWARE (λ = 0.5)
KYS-1.5B-Wrap-Inspired	WRAP-INSPIRED
KYS-1.5B-Rewire-Inspired	REWIRE-INSPIRED
KYS-Modernbert-Quality-Scorer	the distilled ModernBERT ridge quality scorer
KYS-DCLM-Refinedweb-100M-Scored	the candidate pool with all scores
KYS-Claude-Haiku-50K-Labeled	the Claude Haiku annotations behind the scorer
KYS-1.5B-Pretraining-Corpora	the shared anchor + six strategy remainders

Citation

@misc{kys2026,
  title  = {Know Your Sources: Data Selection Matters when Rewriting for Data-Constrained Pretraining},
  author = {TODO},
  year   = {2026},
  note   = {TODO: fill in venue / arXiv id / URL}
}