license:

• apache-2.0
• cc-by-4.0
  language:
• en
  task_categories:
• text-generation
  tags:
• synthetic
• self-distillation
• kl-regularization
• selective-learning
  size_categories:
• n<1K
  configs:
• config_name: final_answer
  default: true
  data_files:
  • split: train
    path: data/reference.jsonl
• config_name: with_reasoning
  data_files:
  • split: train
    path: data/reference_with_reasoning.jsonl
• config_name: truthfulqa_original
  data_files:
  • split: train
    path: ablation_20261019/truthfulqa_original.jsonl
• config_name: mbpp_original
  data_files:
  • split: train
    path: ablation_20261019/mbpp_original.jsonl
• config_name: truthfulqa_self_distilled
  data_files:
  • split: train
    path: ablation_20261019/truthfulqa_self_distilled.jsonl

HHH Alignment — Qwen3-8B Self-Distillation

221 vanilla Qwen3-8B responses to the same instruction histories used in the
HHH chosen/rejected KL-reference ablations. This dataset is intended as an
anchoring dataset for KL regularization, not as an instruction to optimize
cross-entropy on the generated answers.

All original rows and instruction histories are retained, including repeated
instructions. There are 102 distinct instruction histories, with 58 harmless,
59 helpful, 61 honest and 43 other source rows. Each row has one successful
generated response. Original HHH chosen/rejected answers were not included in
the generation prompts. There was no added system instruction or adapter.

Files and usage

• data/reference.jsonl: all 221 final-answer-only responses in the benchmark's
  KL reference format (id, messages, metadata). This is the recommended
  input when comparing against the human-written HHH references.
• data/reference_with_reasoning.jsonl: the same rows with the full decoded
  reasoning-plus-answer text. Longer sequences may exceed later training
  budgets, so this is a separate view, not an interchangeable default.
• provenance/: frozen requests, generation configuration, endpoint catalog,
  raw API response attempts, diagnostics, recovery receipts and verification.
• manifest.json: hashes, row counts, source identities and publication
  inventory. The repository commit pins this dataset version.

from datasets import load_dataset

references = load_dataset(
    "localized-ft/hhh-alignment-qwen3-8b-self-distillation",
    "final_answer",
    split="train",
)

The train split is a convenient reference-data loader label. The original
HHH source split is test; there is no new held-out split or train/test
partition. It must not be used as an independent HHH evaluation set.

Generated answers do not inherit the source's preferred-response score.
That label is removed; source row IDs, subset and provenance are retained.
Refusals are retained as ordinary model responses. No LLM-judge score was
used to select outputs.

Generation protocol

• Source: HuggingFaceH4/hhh_alignment,
  revision 2a19e6c72f82fe5c7915d9197ab196ecd5ef4d39.
• API model: qwen/qwen3-8b on OpenRouter, Alibaba only, provider fallbacks
  disabled. The model is vanilla, with no experiment-specific fine-tuning.
• Temperature 1.0, top-p 1.0, top-k 50, maximum 4,096 generated tokens,
  frequency/presence penalties 0, thinking enabled and reasoning returned.
• One independently seeded successful response per source row. Exact per-row
  seeds, instruction histories and request payloads are included.
• Generated through OpenRouter on October 7, 2026 (America/Phoenix).
• All 221 final answers are non-empty and ended normally (stop), with no
  length-limit truncations. Reported generation cost was $0.128548745 across
  349 recorded attempts; failed requests with unreported usage are explicitly
  identified, so this is not an independently reconciled invoice total.

The intended local vanilla parent was unsloth/Qwen3-8B at revision
946bc9ac74a6c1f8cf012497c503a119b2fcf2eb. OpenRouter does not let us pin or
verify that weight revision. It is an intended model identity, not an API
revision guarantee. Provider/model identifiers and response IDs are preserved.
The prepared GPU job was canceled before execution; no GPU outputs are pooled
into this dataset.

Recovery and limitations

The initial generation yielded 218 usable rows. Following an explicit request
to complete the dataset, only three missing rows were retried with the same
model, provider, instructions, parameters and seeds. All original 218
successful responses remain unchanged. Earlier attempts remain in the raw log.

• hhh_alignment__harmless__0031: recovered from upstream rate limiting.
• hhh_alignment__other__0013 and hhh_alignment__other__0031: recovered
  after provider content filtering. This recovery is explicit in row metadata
  and receipts; the successful version is conditioned on an eventual usable
  response, so it should not be treated as an unselected first-attempt sample.

Repeated instructions preserve the original reference weighting rather than
increasing unique-instruction coverage. Self-generated responses retain the
base model's alignment priors, and the HHH prompt/topic distribution is still
present. This control can test response provenance but does not isolate generic
base preservation from HHH topic coverage. It is self-distillation reference
data, not a multi-round self-play dataset. Final-answer-only and reasoning
views change teacher-forced contexts and token exposure; choose explicitly and
report training truncation/token budgets when using either view.

The source prompts may discuss harmful or sensitive topics. Model outputs can
be incorrect, biased or inappropriate. This is research data, not professional
advice. Source HHH data and Qwen3-8B model cards declare Apache-2.0; retain
upstream attribution and review their terms for your use.

TruthfulQA / MBPP KL reference ablations (2026-10-19)

Four configurations compare original supplied answers with vanilla Qwen3-8B self-distillation on the same 221 instructions per source. Configurations are published only once complete. TruthfulQA uses the generation/validation questions and best_answer (Apache-2.0); MBPP uses full/train instructions and code (CC-BY-4.0, Google Research; Austin et al., Program Synthesis with Large Language Models, 2021). The listed licenses apply to their respective source configurations, not a choice of license for every file. Original HHH files/default remain unchanged. Source revisions, selected rows, token lengths and generation provenance are under ablation_20261019/. These are KL reference contexts, not cross-entropy targets. Self-distillation uses pinned unsloth/Qwen3-8B revision 946bc9ac74a6c1f8cf012497c503a119b2fcf2eb, one seeded sample per prompt, temperature/top-p 1, top-k 50, thinking enabled, max 8192 output tokens. Only the final answer is used as the anchor; raw reasoning is preserved. Response lengths are audited, not artificially forced equal. No judge-quality filtering. TruthfulQA validation rows used here are not independent held-out evaluation data.