Crownelius/GPT-5.6-Sol-Luna-Terra-Traces

GPT-5.6 — Sol · Terra · Luna Library

A maintained mirror of every GPT-5.6 Sol / Terra / Luna dataset on Hugging Face — content-verified, attributed, in one place.

Dataset Viewer | Parquet

// what this is

This is a maintained library — a community mirror of every publicly-available GPT-5.6 Sol / Terra / Luna dataset on Hugging Face, aggregated, validity-filtered, and content-verified with per-row source attribution. It is not Crownelius' own data. Every row credits its original uploader in first_source_dataset, and each contributing author is cited below. New verified sources — and new uploads to existing sources — are pulled in as they appear.

The three variants

Variant	Rows	Source (original author)	Notes
🟠 Sol	14,467	greghavens/gpt-5.6-sol-coding-and-debugging-traces	Codex-CLI agentic coding & debugging rollouts; teacher_model: gpt-5.6-sol, xhigh reasoning, acceptance-test verified.
🟢 Terra	0	—	No public GPT-5.6-Terra dataset exists on Hugging Face yet. Terra is named here as intended scope; verified Terra sources will be added the moment they appear.
🔵 Luna	886	empero-ai/gpt-5.6-luna-sft-900x	Diverse synthetic SFT distilled from openai/gpt-5.6-luna. Note: the assistant is given a "Qwythos / Empero AI" persona — genuine gpt-5.6-luna output, but a persona-injected distill, not raw traces.

Total: 15,353 content-verified rows.

Changelog

• 2026-07-28 — Refreshed from upstream: +9,065 new Sol rows as greghavens expanded their dataset (8,697 re-serialized rows were correctly rejected as payload-duplicates). Sol 5,402 → 14,467; library 6,288 → 15,353.
• 2026-07-16 — Library created: Sol (greghavens) + Luna (Empero AI), Terra listed at 0.

Provenance & honesty

Every retained row is content-verified as genuine GPT-5.6 output — Sol rows carry teacher_model: gpt-5.6-sol + Codex call_… tool-IDs; Luna rows carry model: openai/gpt-5.6-luna. Neither carries Anthropic (toolu_/claude) or foreign-model signals. This provenance is source-asserted and content-verified — it is not, and cannot be, cryptographically certified by OpenAI, and this dataset claims no OpenAI endorsement. Terra is listed at 0 rows rather than filled with anything unverified — a title should not promise data that isn't here.

Rows are deduplicated by both sha256(row_json) and a payload fingerprint (content hashed independently of metadata wrappers), so re-serialized or re-chunked copies of rows already held cannot re-enter the library.

Sibling library (Anthropic models): Crownelius/Complete-FABLE.5-traces-2M. Kept in separate repos so each holds one honest model lineage.

🌐  The Trace Atlas — how these model families actually differ _{(a Fable-authored comparison across all five libraries · click to expand)}

// written by fable · trace atlas v1

Five libraries now mirror five different lineages. They are not interchangeable corpora — each family leaves a different fingerprint in its traces, and if you train on them as if they were the same thing, you inherit the wrong habits. Here is what actually separates them.

ANTHROPIC

Fable 5 · Opus · Sonnet

Signature: the trace is the work. Long agentic sessions where the model reads files, runs commands, reads the error, and revises — toolu_… IDs, tool_result envelopes, self-correction mid-session. Opus/Sonnet rows are the opposite shape: single-turn, dense reasoning with no tools.

Use for: agentic coding, tool-use discipline, long-horizon recovery.

OPENAI

GPT-5.6 Sol · Luna

Signature: verifier-shaped. Sol rows are Codex-CLI rollouts carrying reasoning_effort: xhigh and an explicit verifier: acceptance-tests+quality-review — the work was graded before it was published. Luna is a different animal: persona-injected chat distillation, not raw capture.

Use for: test-passing code, security review, graded outcomes.

ZHIPU AI

GLM-5.2

Signature: the most subject-partitioned corpus here — conversation, science and logic-puzzle sets arrive as separate, deliberately-built slices rather than one undifferentiated dump. Nearly all of it is explicit chain-of-thought; very little is agentic.

Use for: structured CoT, science QA, formal logic.

ALIBABA · TONGYI

Qwen

Signature: the widest version spread — Qwen3, 3.5 and 3.8-Max sit side by side, so the same prompt style appears at several capability levels. Heavily distillation-oriented, and the most internally duplicated ecosystem we measured (one set was an exact 2× copy of itself).

Use for: distillation baselines, cross-version comparison.

MOONSHOT AI

Kimi K3

Signature: scarce and new. Only two genuine public K3 datasets exist, so this is the smallest library by an order of magnitude — an honest snapshot of a frontier that hasn't been mirrored yet, not a shortfall in curation.

Use for: early K3 signal; treat as a seed, not a corpus.

The three axes that actually matter

Axis	What it separates	Where each family sits
Capture vs. distillation	Was the session recorded, or was it re-generated?	Fable 5 & Sol are captured (harness transcripts). Opus, Sonnet, GLM, Qwen, Luna are distilled — cleaner, but one step removed from real behaviour.
Agentic vs. single-turn	Does the model act, or just answer?	Fable 5 / Sol / Kimi K3 carry tool calls and their results. Opus, Sonnet, GLM and Qwen are overwhelmingly one prompt → one long reasoned answer.
Verifiability of origin	Can you tell which model wrote it?	Strongest where rows carry an in-row model / teacher_model field (greghavens sets, Roman1111111, r0b0tlab). Weakest for reasoning distills, which never self-identify — those are card-asserted only, and labelled as such.

⚠ the trap these libraries exist to avoid

Cross-model contamination is rampant and almost always invisible from the title. Verified examples caught while building these: a GLM-5.2 set carrying Anthropic toolu_ IDs; a "Claude Sonnet" set whose reasoning was rewritten by Gemma-4-31B; a claude-fable-5-tagged set that was actually MiniMax MiMo; a Sonnet corpus silently blended with Gemini 3.1 Pro; and a "Fable5" dataset that turned out to be Sumerian cuneiform OCR. Every library here lists what it rejected, and why.

Claude · GPT-5.6 · GLM-5.2 · Qwen · Kimi K3

Schema

Column	Type	Meaning
row_hash	string	sha256(row_json) — stable de-duplication key
first_source_dataset	string	Upstream dataset / original author (also identifies the variant)
first_source_config / first_source_split	string	Upstream config / split
first_source_row_index	int64	Index within the upstream source
seen_count	int64	Times this canonical row was seen during aggregation
row_json	string	The full source row as JSON — parse for messages, tools, model / teacher_model, etc.

Citations & attribution

All content belongs to its original uploaders and is re-hosted under their licenses, with full credit:

Rows	Original author / dataset	License
14,467	greghavens — gpt-5.6-sol-coding-and-debugging-traces	CC-BY-4.0
886	Empero AI — gpt-5.6-luna-sft-900x	see upstream card

If you use this library, please cite the original authors above — not this mirror.

Loading

from datasets import load_dataset
ds = load_dataset("Crownelius/GPT-5.6-Sol-Luna-Terra-Traces", split="train")

# filter to a variant via provenance:
sol  = ds.filter(lambda r: "gpt-5.6-sol"  in r["first_source_dataset"])
luna = ds.filter(lambda r: "gpt-5.6-luna" in r["first_source_dataset"])