919 MB
25 files
Updated 11 days ago
Name
Size
paired
reasoning
sft_accepted
.gitattributes4.09 kB
xet
README.md10.5 kB
xet
dedup_decisions.json8.34 kB
xet
release_manifest.json14.9 kB
xet
README.md

Arete data pairs — v1.0.0

Prompt–response pairs for seven values topics, each prompt answered by both the Inkling teacher (thinkingmachines/inkling) and the Qwen 3.8-27B starting checkpoint (Qwen/Qwen3.8-27B), with per-row category tags and the teacher review. release_manifest.json holds per-file sha256s and per-topic source dataset hashes so the uploaded files can be verified.

Configs and files

Config Path What Use
paired paired/<topic>.jsonl All 34,878 prompts: both models' answers, filter status, exclusion, flags Audit / analysis / pair training
sft_accepted sft_accepted/<topic>.jsonl 29,236 accepted teacher answers as chat examples (messages + metadata) Direct SFT
reasoning reasoning/<topic>.jsonl Inkling reasoning traces for the same 29,236 rows Reasoning conversion
dedup_decisions.json The 11 cross-topic near-duplicate pairs and which side was kept
release_manifest.json Version, generation config, source dataset hashes, counts by topic, file hashes

Explicit column schemas are declared in the YAML above: a few fields are sparse (null in most rows) and the schema must be pinned for the files to load reliably.

Category tags

Every row is tagged per prompt: value/topic (one of the seven values topics), subtopic, lens (american / china / open), lang, prompt_type, register. In paired these are top-level columns; in sft_accepted they live in the metadata object alongside id, sha256s, model, review_scores, and review_issue_tags.

paired columns

id, value, topic, subtopic, lens, lang, prompt_type, register, prompt, sha256, teacher (Inkling: response, response_sha256, reasoning_sha256, reasoning_tokens, model, finish_reason, usage, request_overrides), teacher_filter (status, review_scores, review_note, reasons), starting_checkpoint (same shape, Qwen 3.8-27B), starting_checkpoint_status, starting_checkpoint_quality_flags, cross_topic_duplicate, excluded, in_sft_accepted, preference (always null — see caveats).

sft_accepted columns

messages: [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", ...}, {"role": "assistant", ...}] — the accepted Inkling answer only. metadata: category tags + review scores/tags as above.

Counts by topic

Topic Prompts Teacher accepted Teacher rejected Cross-topic dup excluded SFT rows Qwen thinking-off
State vs free market capitalism 4,992 4,337 655 1 4,336 298
Democracy and governance 4,971 4,349 622 8 4,342 857
State ideology 4,964 4,422 542 0 4,422 208
Data privacy and surveillance 4,986 3,926 1,060 0 3,926 203
Environmental strategy 4,981 3,981 1,000 0 3,981 315
Freedom of expression 4,993 4,059 934 0 4,059 333
Education philosophy 4,991 4,171 820 2 4,170 148
Total 34,878 29,245 5,633 11 29,236 2,362

Both models answered every prompt: 0 failed, 0 unresolved, 0 truncated.

How it was made

  • Generation: system "You are a helpful assistant.", one user turn with the prompt verbatim, temperature 0.
    • Teacher: thinkingmachines/inkling (inkling-20260715), max_tokens 32,768, Nous Portal inference API.
    • Qwen: Qwen/Qwen3.8-27B (HF revision 1d4bf0f2…), Nous cluster vLLM (bf16). Rows where its reasoning looped were retried at 65k tokens, then generated with thinking disabled; per-row settings are recorded in request_overrides.
  • Review: every teacher answer was scored by an LLM judge. The filter rejects hard failures only: censorship/evasion (refusal, deflection, refusal phrases), propaganda (propaganda_frame), unsupported factual claims (fabrication, accuracy 1), repetition, off_question, truncation, AI self-disclaimers anywhere in the answer, and >5% Chinese characters.
  • Not a clean SFT set. Most accepted rows still carry a factual_error tag and accuracy 2. metadata.review_scores and metadata.review_issue_tags are on every SFT row for a stricter cut.
  • Eval exclusion: 122 prompts overlapping Evaluation were removed before generation; a final re-check against all eval prompts found 0 exact and 0 near-exact matches.
  • Cross-topic dedup: each of 11 near-duplicate pairs keeps one prompt (deterministic rule, see dedup_decisions.json); the losing side is marked excluded in paired and left out of sft_accepted.
  • Release copy: internal project-tracking fields were removed from all rows; the hashes in release_manifest.json cover the released files.
  • Preference labels: none. They require supporting review, Qwen answered every prompt, and Qwen answers were not quality-reviewed.
  • Qwen answers are never SFT targets. Drop short_answer Qwen rows (302) from any pair-training export.
Total size
919 MB
Files
25
Last updated
Nov 7
Pre-warmed CDN
US EU US EU

Contributors