Solstice-AI/Project-Solace-1.0-GLM5.2-Fable5-Opus5-DeepseekV4-Pro0813-Preview

Project Solace

The largest verified frontier-model distillation corpus ever released.

60 datasets · 7 frontier model families · 12,586,893 unique conversations · One file · Zero filler

---

The short version

This is synthetic data. The best kind of synthetic data. Every example was generated by a verified 2026 frontier model — GLM-5.2, Claude Fable 5, Mythos 5, GPT-5.6 Sol, GPT-5.5 Codex, DeepSeek V4 Pro 0813, Qwen 3.8-Max, and Kimi K3 — then exact-deduplicated and intelligently shuffled into a single drop-in JSONL. In 2026, frontier distillation is the gold standard for post-training. This is a pure vein of it. Load it, train it, done.

---

Why Synthetic Data Won

Distillation is how the frontier gets taught. The strongest open models in 2026 are built on carefully selected outputs from stronger models — real traces, real trajectories, real reasoning, captured verbatim and replayed at scale. Not regenerated by smaller models, not paraphrased into mediocrity. The original output of the best models that exist.

Project Solace is that practice taken seriously. It aggregates the strongest frontier-model traces the open-source community has produced — agentic rollouts, coding and debugging sessions, ARC-AGI3 reasoning, SWE-bench replays, math and STEM reasoning — into one corpus where every row is accounted for, every byte is real model output, and nothing is filler.

"The best training data isn't generated by you. It's distilled from the models that matter."

---

The Numbers (measured, not estimated)

Metric	Value
Unique conversations	12,586,893
Source rows ingested	16,032,427
Exact duplicates removed	3,445,534 (21.5% — cross-dataset overlap)
Uncompressed size	137.7 GB
Compressed file	solace_final.jsonl.gz — 34.3 GB
Source datasets	60 (7 frontier model families)
Format	JSONL, OpenAI messages format
Deduplication	Exact — SHA256 of normalized messages JSON
Ordering	Intelligently shuffled (weighted round-robin)
Date range	May – August 2026

Every row is tagged with its source dataset and originating model — filter, rebalance, or harvest any class of example in one line.

---

What Makes This Different

Typical "distillation" dump	Project Solace
Outputs from small or unknown models	Verified 2026 frontier models only
Unverified claims about provenance	Every dataset manually audited, every file accounted for
Legacy models mixed in	Not one token from a pre-2026 model
No deduplication	Exact SHA256 dedup — within AND across all 60 sources
Sequential dump order	Intelligently shuffled — proportional representation in every window
Generic chat filler	Agentic traces, coding, debugging, ARC-AGI3, math, SWE-bench
Many files to manage	One JSONL file. Done.

---

What's Inside

7 frontier model families. Not one token from a legacy model.

Model	Focus	Highlights
GLM-5.2	Agent rollouts, coding, reasoning	The dominant open-source frontier — OpenHands rollouts, on-policy regeneration, ARC traces
Claude Fable 5	Agentic coding, SFT	Anthropic's latest reasoning engine — agentic traces you can't get anywhere else
GPT-5.6 Sol	Coding, debugging, ARC-AGI3	OpenAI's frontier — coding traces + abstract reasoning challenges
GPT-5.5 Codex	ARC-AGI3, code generation	The original code-specialized frontier model
DeepSeek V4 Pro 0813	Math, STEM, SWE-bench, reasoning	200K-distilled math/STEM, ResearchMath, real SWE-bench replays
Qwen 3.8-Max	Multi-teacher distillation, coding	Cross-architecture distillation data
Kimi K3 / Opus 4.7 / Multi	Reasoning, SWE agent, multi-teacher	Moonshot + Anthropic reasoning specialists + 6-model cross-distillation

---

Data Format

Every example is a single JSON line — no transformation, no preprocessing, just load and train:

{"messages":[{"role":"system","content":"..."},{"role":"user","content":"..."},{"role":"assistant","content":"..."}],"source":"repo/dataset-name","model":"glm-5.2"}

Every example is tagged with its source dataset and originating model. Tool calls, reasoning content, and full multi-turn agent histories are preserved verbatim — nothing is dropped, nothing is rewritten.

Compatible with everything

Framework	Status
OpenAI Fine-tuning API	Drop-in
HuggingFace TRL / SFTTrainer	Drop-in
Axolotl	Drop-in
LLaMA-Factory	Drop-in
torchtune	Drop-in
Any JSONL reader	Works

---

Quickstart

1. Decompress (one-time)

# ~34 GB compressed → 137.7 GB decompressed
zcat solace_final.jsonl.gz > solace_final.jsonl

2. Load and train in 3 lines

from datasets import load_dataset
from trl import SFTTrainer, SFTConfig

dataset = load_dataset(
    "Solstice-AI/Project-Solace-1.0-GLM5.2-Fable5-Opus5-DeepseekV4-Pro0813-Preview",
    data_files="solace_final.jsonl.gz", split="train"
)
trainer = SFTTrainer(model=model, train_dataset=dataset, args=SFTConfig(output_dir="./output"))
trainer.train()

That's it. No preprocessing. No formatting. Just load and go.

3. Filter by model — train on what matters

import json

with open("solace_final.jsonl") as f:
    examples = [json.loads(l) for l in f]

# GLM-5.2 only — agent rollouts and coding
glm = [e for e in examples if e["model"] == "glm-5.2"]

# Fable 5 only — agentic coding traces
fable = [e for e in examples if e["model"] == "fable5"]

# DeepSeek V4 Pro only — math + STEM
deepseek = [e for e in examples if e["model"] == "deepseek-v4"]

4. Filter by capability

# All coding-focused data
coding = [e for e in examples if any(k in json.dumps(e).lower()
         for k in ["code", "debug", "programming", "swe"])]

# All reasoning-focused data
reasoning = [e for e in examples if any(k in json.dumps(e).lower()
            for k in ["reason", "proof", "math", "logic"])]

---

How This Was Built

This wasn't a weekend project. This was a systematic, surgical extraction of the strongest training data the open-source community has produced.

Phase 1 — Discovery

Searched HuggingFace across hundreds of queries — every frontier model name, every distillation keyword, every agentic-trace pattern. Found 70+ candidate datasets.

Phase 2 — Verification

Every dataset was inspected by hand, file by file:

• Is this actual frontier-model output? (Not weights, not benchmarks)
• Is this from a 2026 frontier model? (No legacy models sneaking in)
• Is this coding / reasoning / agentic? (Not generic chat filler)
• Is every file accounted for? (Parquet, JSONL, CSV, tar.zst — all of it, nothing skipped)

Rejected: cuneiform tablet translations, SOC logs, weight files, calibration data, and dozens of pre-2026 model dumps.

Phase 3 — Merge

A universal format processor decoded every schema family — OpenAI messages, conversations (from/value), context/completion, Alpaca instruction/input/output, Harmony-renderer transcripts, ARC-AGI3 agent trajectories, Claude-Code session logs, Codex response_item streams, and more. Tool calls and reasoning content preserved verbatim. Every character kept.

Phase 4 — Dedup

Exact deduplication only. SHA256 of the normalized messages JSON — byte-identical examples are removed, within and across all 60 source datasets. No fuzzy matching, no semantic clustering. If two examples differ in any way, they both stay. This removed 3.4M redundant copies (21.5%) — the same conversations genuinely shipping in multiple upstream datasets.

Phase 5 — Intelligent shuffle

Instead of a naive sequential dump, the corpus is weighted round-robin interleaved: every window of ~100–1000 rows contains proportional representation from every source. No dataset dominates a stretch; no model disappears for long. Your model sees the full distribution at every point of training — while the source tag on every row still lets you harvest any class you want.

Phase 6 — Verification

The final file was verified end-to-end — full-stream re-hash of all 12,586,893 rows, zero duplicates, zero malformed rows, and a byte-exact SHA256 check against the uploaded artifact.

---

Datasets Included (60)

Click to expand the full source list — every dataset credited to its creator below.

GLM-5.2 (15 datasets)

Dataset	Type
OnepointfiveHz/glm-5.2-openhands-rollout-0806	Agent rollouts
mgoin/open-perfectblend-glm5.2-regen	On-policy regeneration
OnepointfiveHz/glm-5.2-openhands-rollout-0805	Agent rollouts
OnepointfiveHz/glm-5.2-openhands-rollout-0801_07	Agent rollouts
JessieWei/GLM-5.2-FP8-nemotron-codealpaca-thinking	Thinking traces
mgoin/GLM-5.2-FP8-magpie-ultrachat	Reasoning
DavidrPatton/Fable-5-GLM-5.2-Traces	Dual teacher
ansulev/GLM-5.2-Conversation	Conversation
AgentNativeResearchLab/arc-agi3-cc-glm5.2-ls20	ARC traces
AgentNativeResearchLab/arc-agi3-cc-glm5.2-su15	ARC traces
AletheiaResearch/GLM-5.2-Agent	Agent
greghavens/glm-5.2-coding-and-debugging-traces	Coding
marin-community/openthoughts4-code-9168-prompts-glm-5.2-n4	Code reasoning
AgentNativeResearchLab/fm-open-problems-glm5.2-trajectories	Trajectories
AgentNativeResearchLab/lhtb-glm5.2-trajectories	Trajectories

Claude Fable 5 (12 datasets)

Dataset	Type
usernamebetter/fable5-traces-agentic	Agentic
usernamebetter/fable5-traces-agentic-clean	Agentic (cleaned)
usernamebetter/fable5-traces-agentic-clean-v2	Agentic v2
saidutta69/fable-5-premium	Premium
Manusagents/Vibe-Coding-Claude-Fable-5	Vibe coding
Biomechanist/fable-5-coding-and-debugging-traces-harmonized	Harmonized
AgentNativeResearchLab/arc-agi3-cc-fable5-ls20	ARC traces
AgentNativeResearchLab/arc-agi3-cc-fable5-g50t	ARC traces
Solstice-AI/Complete-FABLE.5-traces-2M	Complete traces
aisamdasu/algocean-fable5-traces	Coding
Nexlab/fable5-agentic-coding-sft	Agentic SFT
Swarm-AI-Research/fable5-traces-sft	SFT

GPT-5.6 Sol (7 datasets)

Dataset	Type
AgentNativeResearchLab/arc-agi3-codex-gpt5.6sol-g50t	ARC reasoning
greghavens/gpt-5.6-sol-coding-and-debugging-traces	Coding
Biomechanist/gpt-5.6-sol-coding-and-debugging-traces-harmonized	Harmonized
AgentNativeResearchLab/arc-agi3-codex-gpt5.6sol-r11l	ARC traces
AgentNativeResearchLab/arc-agi3-codex-gpt5.6sol-ls20	ARC traces
AgentNativeResearchLab/arc-agi3-codex-gpt5.6sol-su15	ARC traces
AgentNativeResearchLab/fm-open-problems-gpt5.6-sol-trajectories	Trajectories

GPT-5.5 Codex (7 datasets)

Dataset	Type
AgentNativeResearchLab/arc-agi3-codex-gpt5.5-g50t	ARC reasoning
AgentNativeResearchLab/arc-agi3-codex-gpt5.5-s5i5	ARC traces
AgentNativeResearchLab/arc-agi3-codex-gpt5.5-r11l	ARC traces
AgentNativeResearchLab/arc-agi3-codex-gpt5.5-ls20	ARC traces
AgentNativeResearchLab/arc-agi3-codex-gpt5.5-ar25	ARC traces
AgentNativeResearchLab/arc-agi3-codex-gpt5.5-su15	ARC traces
AletheiaResearch/GPT-5.5-Codex	Agent traces

DeepSeek V4 Pro 0813 (9 datasets)

Dataset	Type
sequelbox/Mitakihara2-DeepSeek-V4-Pro	Coding
ansulev/deepseek-v4-pro-tachibana4	Coding
sequelbox/Titanium4-DeepSeek-V4-Pro	Coding
trjxter/DeepSeek-V4-Pro-Reasoning-8000x	Reasoning
ansulev/deepseek-v4-pro-agent	Agent
Jackrong/DeepSeek-V4-Pro-Distilled-200K	Math/STEM
Cartinoe5930/ResearchMath-14k_deepseek-v4-pro	Math
r0b0tlab/deepseek-v4-pro-0813-agentic	Agentic
fxiao0369/deepseek-v4-pro-swebench-replay	SWE-bench

Qwen 3.8-Max (3 datasets)

Dataset	Type
CodeFlame/Qwen3.8-GLM5.2-Kimi-K3-GPT5.6-Gemini-3.1-Claude-Fable5-Mythos5-distillation	Multi-teacher
r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation	Multi-teacher
guell00/qwen-3.8-code	Code

Kimi K3 / Opus 4.7 / Multi (6 datasets)

Dataset	Type
lzy510016411/fable5-gpt5.5-opus4.7-mixed-agent-traces	Multi-model
Hein1212/claude-opus-4.6-4.7-reasoning-8.7k	Reasoning
Manusagents/Mythos-5-and-Fabel-5-Class-Model-Outputs	Multi-model
AgentNativeResearchLab/fm-open-problems-kimi-k3-trajectories	Trajectories
thientrangngv/SERA-KimiK3-Django-SWEAgent-Raw-T1	SWE agent
thientrangngv/SERA-KimiK3-Django-SWEAgent-Cliff32k-T1	SWE agent

Manusagents (1 dataset)

Dataset	Type
Manusagents/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-5-Mythos-5-Qwen-3.7-Max-and-more-Distillation-Dataset	Multi-teacher

---

Duplicates Removed

A handful of upstream datasets were found to be duplicates of others in the corpus and were dropped entirely:

Removed	Duplicate of	Reason
JessieWei/GLM-5.2-FP8-nemotron-codealpaca (non-thinking)	…-thinking sibling	Same content, no reasoning
turintech/GLM-5.2-MaxMin108-saliency	—	Artifact-only (interpretability tables, zero model output)
JacobChang/GLM5.2-B200-profiles	—	Artifact-only (GPU profiling, zero model output)
RadixArk/Qwen3.8-27B-Regen-Mixture-v1	—	Unavailable (404)

The remaining 60 datasets were merged, then exact-deduplicated (SHA256 of the messages JSON) — 3,445,534 byte-identical cross-dataset copies removed. Zero fuzzy matching.

---

Manifest

See manifest.jsonl for per-dataset statistics: rows, characters, token estimates, and source links.

---

Citation

@dataset{solace2026,
  title={Project Solace: Frontier Model Distillation Corpus},
  author={Solstice-AI},
  year={2026},
  url={https://huggingface.co/datasets/Solstice-AI/Project-Solace-1.0-GLM5.2-Fable5-Opus5-DeepseekV4-Pro0813-Preview},
  note={60 datasets, 12,586,893 unique conversations, exact-deduplicated and intelligently shuffled}
}

---

Why "Solace"?

solace (noun): comfort in a time of distress or disappointment.

The open-source community has been disappointed by data quality for too long. Disorganized or low-quality data that makes models worse. Legacy outputs that teach bad habits. Datasets that are 90% metadata and 10% useful text.

Project Solace is the remedy. Every token was produced by a model that actually matters. Every example was verified. Every file was accounted for. No shortcuts. No filler. Just the strongest distillation corpus the community has access to.

---

Acknowledgments

This corpus aggregates data from 60 datasets created by researchers across the open-source AI community. This project would not exist without their work, and full credit belongs to every upstream author:

Creator	Datasets
AgentNativeResearchLab	18 datasets — ARC-AGI3 trajectories across GLM-5.2, Fable 5, GPT-5.6 Sol, GPT-5.5 Codex; FM open-problems trajectories; LHTB trajectories; Kimi K3 trajectories
OnepointfiveHz	3 datasets — GLM-5.2 OpenHands agent rollouts
usernamebetter	3 datasets — Fable 5 agentic traces (raw, clean, v2)
mgoin	2 datasets — GLM-5.2 on-policy regeneration, Magpie reasoning
Biomechanist	2 datasets — harmonized coding/debugging traces (Fable 5, GPT-5.6 Sol)
greghavens	2 datasets — coding and debugging traces (GLM-5.2, GPT-5.6 Sol)
AletheiaResearch	2 datasets — GLM-5.2 agent, GPT-5.5 Codex traces
ansulev	3 datasets — GLM-5.2 conversation, DeepSeek V4 Pro coding, DeepSeek V4 Pro agent
r0b0tlab	2 datasets — DeepSeek V4 Pro 0813 agentic, Qwen 3.8-Max multi-teacher
sequelbox	2 datasets — DeepSeek V4 Pro coding (Mitakihara2, Titanium4)
thientrangngv	2 datasets — SERA Kimi K3 Django SWE-agent (Raw, Cliff32k)
Manusagents	3 datasets — Vibe Coding Fable 5, Mythos 5/Fable 5 outputs, 7-model distillation
Solstice-AI	1 dataset — Complete FABLE.5 traces 2M
JessieWei	1 dataset — GLM-5.2 FP8 thinking traces
DavidrPatton	1 dataset — Fable 5 / GLM-5.2 dual-teacher traces
saidutta69	1 dataset — Fable 5 premium
aisamdasu	1 dataset — Alg-ocean Fable 5 traces
Nexlab	1 dataset — Fable 5 agentic coding SFT
Swarm-AI-Research	1 dataset — Fable 5 traces SFT
marin-community	1 dataset — OpenThoughts4 code prompts
trjxter	1 dataset — DeepSeek V4 Pro reasoning 8K
Jackrong	1 dataset — DeepSeek V4 Pro distilled 200K
Cartinoe5930	1 dataset — ResearchMath 14K
fxiao0369	1 dataset — DeepSeek V4 Pro SWE-bench replay
CodeFlame	1 dataset — 7-model distillation
guell00	1 dataset — Qwen 3.8 code
lzy510016411	1 dataset — Fable 5 / GPT-5.5 / Opus 4.7 mixed traces
Hein1212	1 dataset — Claude Opus 4.6/4.7 reasoning

Individual datasets retain their original licenses and terms — please respect each upstream author's license on their own dataset page.

---

License

Individual datasets retain their original licenses. This merged corpus is provided as-is for research purposes. Project Solace uses EPORAUL v17 custom licensing — please read the license documents before downloading.

---

Built by

Solstice-AI — pushing the frontier of open-source AI training data.

---

If this corpus helps your research or project, we'd love to hear about it. Star the repo, drop a citation, or reach out.