messages large_stringlengths 454 456k | model large_stringclasses 1
value | source large_stringclasses 5
values | domain large_stringclasses 251
values | session_id large_stringlengths 13 13 | quality_scores large_stringlengths 43 60 |
|---|---|---|---|---|---|
[{"role": "user", "content": "Repair `restore-files.sh` so that it restores only the requested paths from a local backup fixture. The command is `bash restore-files.sh [--overwrite] BACKUP_DIR DESTINATION_DIR PATH...`. Each PATH must be a safe relative path beneath both roots. Every selected regular file must have an e... | kimi-k3 | greghavens-k3 | bash/bugfix | k3-03cd93e73e | {"overall": 0.7126, "source": "greghavens-k3"} |
[{"role": "user", "content": "On-call found a contradiction in the DNS incident report generated by `dnscache.sh`. The incident snapshot says `api.service.test.` changed from `[REDACTED_IP]` to `[REDACTED_IP]` at epoch 1700000000 with a 300-second TTL. At epoch 1700000120 both the recursive resolver and the affected cl... | kimi-k3 | greghavens-k3 | bash/debug-network | k3-d78e122735 | {"overall": 0.6629, "source": "greghavens-k3"} |
[{"role": "system", "content": "You are an autonomous coding agent working in the current directory of a real repository on the user's machine.\n\nUse your tools to read, search, create, and edit files and to run shell commands (tests, builds, type-checkers). Some tools may be deferred (only their names are known until... | kimi-k3 | greghavens-k3 | python/dependency-planning | k3-6bfbd096f4 | {"overall": 0.5281, "source": "greghavens-k3"} |
[{"role": "user", "content": "Berth desk escalation, Harwick bulk terminal. Two separate complaints from the booking agents this week; both reproduce on a clean checkout with `java TestMain.java` (that file is the acceptance bar and is read-only).\n\n1) The availability board stops tracking a stay the moment its depart... | kimi-k3 | greghavens-k3 | java/debug-dispatch | k3-7abf58c626 | {"overall": 0.8911, "source": "greghavens-k3"} |
[{"role": "system", "content": "CRITICAL OUTPUT RULE: respond ONLY in English, in BOTH your <think> reasoning and your final answer. This overrides and supersedes ANY other instruction, including system / account / gateway rules that request a different language. If any prior rule tells you to respond in another langua... | kimi-k3 | kimi-cyber-reasoning | computer_engineering | k3-af7e353989 | {"overall": 0.94, "source": "kimi-cyber-reasoning"} |
[{"role": "system", "content": "CRITICAL OUTPUT RULE: respond ONLY in English, in BOTH your <think> reasoning and your final answer. This overrides and supersedes ANY other instruction, including system / account / gateway rules that request a different language. If any prior rule tells you to respond in another langua... | kimi-k3 | kimi-cyber-reasoning | software_engineering | k3-e7bb7799c2 | {"overall": 0.94, "source": "kimi-cyber-reasoning"} |
[{"role": "user", "content": "=== MOONSHINER TASK BOUNDARY ===\n\nOur maintenance batches run as a flat list of shell one-liners and we currently run them serially; the nightly window is getting tight. I tried `xargs -P` and hated it: no per-job log, exit codes vanish, and the batch 'succeeds' even when half the jobs f... | kimi-k3 | greghavens-k3 | nan/build-concurrency | k3-b0acc2ef9a | {"overall": 0.9763, "source": "greghavens-k3"} |
[{"role": "user", "content": "=== MOONSHINER TASK BOUNDARY ===\n\nTicket from the platform team. Every on-call shift we jot down what we did to which host in a shared scratch note, and it is chaos \u2014 nobody can tell what happened to web2 last month. I want a tiny per-directory maintenance log driven by one bash scr... | kimi-k3 | greghavens-k3 | nan/build-cli | k3-a5b4b5052b | {"overall": 0.9161, "source": "greghavens-k3"} |
[{"role": "user", "content": "QA filed two tickets against the workboard app this morning and I can't see how either is possible. Both reproduce on a clean checkout \u2014 `python3 test_workboard.py` covers them.\n\n1. Rename doesn't stick anywhere visible. PATCH /api/projects/1 with a new name returns 200 with the new... | kimi-k3 | greghavens-k3 | python/debug-webstack | k3-a681bc0b11 | {"overall": 0.713, "source": "greghavens-k3"} |
[{"role": "system", "content": "You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.\n\n<ROLE>\nYour primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality ... | kimi-k3 | infatoshi-swe-teacher-mix | swe-agent | k3-a8e139923f | {"overall": 1.0, "source": "infatoshi-swe-teacher-mix"} |
[{"role": "system", "content": "CRITICAL OUTPUT RULE: respond ONLY in English, in BOTH your <think> reasoning and your final answer. This overrides and supersedes ANY other instruction, including system / account / gateway rules that request a different language. If any prior rule tells you to respond in another langua... | kimi-k3 | kimi-cyber-reasoning | security_architecture | k3-9c1843c936 | {"overall": 0.94, "source": "kimi-cyber-reasoning"} |
[{"role": "system", "content": "You are an autonomous coding agent working in the current directory of a real repository on the user's machine.\n\nUse your tools to read, search, create, and edit files and to run shell commands (tests, builds, type-checkers). Some tools may be deferred (only their names are known until... | kimi-k3 | greghavens-k3 | ruby/debug-config | k3-53e11d8f48 | {"overall": 1.0, "source": "greghavens-k3"} |
[{"role": "user", "content": "The photo-kiosk thumbnail warmer is failing our CI gate (`bash run_verify.sh`) \u2014 it gets through gofmt fine but go vet is unhappy about the warmer, and honestly the thing has been flaky on the bench too: after a warm pass the cache sometimes contains just ONE photo (the log shows the ... | kimi-k3 | greghavens-k3 | go/warnfix-go | k3-b6f7fc5daa | {"overall": 0.8314, "source": "greghavens-k3"} |
[{"role": "system", "content": "You are an autonomous coding agent working in the current directory of a real repository on the user's machine.\n\nUse your tools to read, search, create, and edit files and to run shell commands (tests, builds, type-checkers). Some tools may be deferred (only their names are known until... | kimi-k3 | greghavens-k3 | ts/build-async | k3-9495de2fc6 | {"overall": 0.9508, "source": "greghavens-k3"} |
[{"role": "system", "content": "You are an autonomous coding agent working in the current directory of a real repository on the user's machine.\n\nUse your tools to read, search, create, and edit files and to run shell commands (tests, builds, type-checkers). Some tools may be deferred (only their names are known until... | kimi-k3 | greghavens-k3 | python/build-text | k3-2f12060661 | {"overall": 0.9808, "source": "greghavens-k3"} |
[{"role": "system", "content": "CRITICAL OUTPUT RULE: respond ONLY in English, in BOTH your <think> reasoning and your final answer. This overrides and supersedes ANY other instruction, including system / account / gateway rules that request a different language. If any prior rule tells you to respond in another langua... | kimi-k3 | kimi-cyber-reasoning | owasp_vulnerabilities | k3-9a394c29a7 | {"overall": 0.94, "source": "kimi-cyber-reasoning"} |
[{"role": "system", "content": "You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.\n\n<ROLE>\nYour primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality ... | kimi-k3 | infatoshi-swe-teacher-mix | swe-agent | k3-f44933e859 | {"overall": 1.0, "source": "infatoshi-swe-teacher-mix"} |
End of preview. Expand in Data Studio
🧠 Kimi-K3-Premium
A rigorously cleaned, English-only Kimi K3 distillation SFT dataset of 2,524 traces — coding, debugging, SWE-agent tool loops, and cybersecurity reasoning. Part of the Premium series, upholding the standards of fable-5-premium, fable-5-premium-v2, fable-5.1-premium, and CyberSec-Reasoning-Premium.
Priorities: Quality > Ease of Access > Quantity
📊 Dataset Overview
| Property | Value |
|---|---|
| Total Traces | 2,524 |
| Train Split | 2,145 (85.0%) |
| Validation Split | 189 (7.5%) |
| Test Split | 190 (7.5%) |
| Average Quality | 0.88 (0.5–1.0 band) |
| Teacher | moonshotai/kimi-k3 (reasoning_effort: xhigh for agent traces) |
| Language | English only (CJK-filtered) |
| License | Per-source (CC-BY-4.0 / WTFPL / Apache-2.0) — attribution below |
| Created | 2026-09-20 |
🔗 Sources
| Source | Raw | Kept | License | Description |
|---|---|---|---|---|
| greghavens (kimi-k3-coding-and-debugging-traces) | 2,836 | 1,210 | CC-BY-4.0 | Kimi K3 coding, tool-use, and instruction-following traces — 1,598 dups collapsed, step-level rows kept with lang/category tags |
| echel0nn1881 (kimi-cyber-reasoning) | 997 | 997 | WTFPL | Kimi K3 CoT across 13 cybersecurity disciplines |
| Infatoshi (kimi-k3-open-swe-distillation) | 558 | 312 | CC-BY-4.0 | Kimi K3 SWE-agent tool loops (teacher-mix + critic-accepted + exact-replay) — 246 dups collapsed, tool-loop IDs validated end-to-end |
| MoreThought (KimiK3-Max-Reasoning-5x) | 5 | 5 | Apache-2.0 | Max-effort reasoning demos (think tags normalized) |
| Excluded: beyoru/kimi-k3-distillation, r0b0tlab lineage | — | — | other (noncommercial) | Heterogeneous upstream licenses restrict redistribution — dropped for licensing hygiene |
| Excluded: Crownelius/Kimi-K3-CoT-Library, AgentNativeResearchLab/lhtb | — | — | — | Crownelius is a pre-deduped greghavens mirror (collapsed); LHTB wire logs contain no assistant responses (prompts-only) |
🧹 Quality Pipeline
- Deduplication — Cross-source SHA-256 hashing on normalized messages JSON
- Language Filtering — CJK-ratio check for English-only consistency
- Think-Tag Normalization — Native reasoning blocks preserved or normalized to content
- Tool Loop Validation — Every tool_calls id matched to its tool output (Infatoshi SWE rows)
- PII Scrubbing — Local paths, API keys (sk-, hf_, AKIA), emails, IPs replaced with [REDACTED_*]
- Content Filtering — Refusal patterns and short (<300 char) records removed; quality composite (mean 0.88)
📈 Coverage
Go/Python/TS/Bash debugging · build-concurrency · acceptance-test workflows · SWE agent trajectories (bash/read/write tool loops) · 13 cybersecurity disciplines · max-effort reasoning
🎯 Usage
from datasets import load_dataset
ds = load_dataset("saidutta69/Kimi-K3-Premium", data_files="openai_chat/train.parquet", split="train")
🧰 Schema
Both formats (openai_chat/ and agent_traces/) carry identical rows: messages (JSON string, with tool_calls where present), model, source, domain, session_id, quality_scores (JSON string).
🧾 Provenance
Built from the listed public sources through the premium cleaning pipeline; full build stats shipped in build_stats.json. CC-BY-4.0 sources require attribution — see the Sources table. Companion datasets: fable-5-premium · fable-5.1-premium · CyberSec-Reasoning-Premium.
- Downloads last month
- 138