Datasets:

saidutta69
/

Qwen3.8-Agent-Premium

messages large_stringlengths 7.17k 538k	model large_stringclasses 1 value	source large_stringclasses 2 values	domain large_stringclasses 16 values	session_id large_stringlengths 14 14	quality_scores large_stringclasses 41 values
[{"role": "user", "content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON with the foll...	qwen3.8	ukisai-qwen38-agent	terminal-nl	q38-1755b2933d	{"overall": 1.0, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-bugs	q38-4702fb682f	{"overall": 1.0, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-bugs	q38-b3009465e6	{"overall": 1.0, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-bugs	q38-3baa79444d	{"overall": 1.0, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-nl	q38-aa92e8eb04	{"overall": 0.92, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-nl	q38-16520fc256	{"overall": 0.88, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-nl	q38-4c6cc8a9b9	{"overall": 0.92, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-bugs	q38-fb53f04aa1	{"overall": 1.0, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-bugs	q38-743286c7f7	{"overall": 0.96, "source": "ukisai-qwen38-agent"}
"[{\"role\": \"user\", \"content\": \"You are an AI assistant tasked with solving command-line tasks(...TRUNCATED)	qwen3.8	ukisai-qwen38-agent	terminal-nl	q38-f44fdf2a2a	{"overall": 0.96, "source": "ukisai-qwen38-agent"}

End of preview. Expand in Data Studio

🤖 Qwen3.8-Agent-Premium

A rigorously cleaned, English-only Qwen3.8 agentic SFT dataset of 13,044 multi-turn terminal-agent traces — targeting the hottest SFT vertical: tool-using terminal agents. Part of the Premium series, upholding the standards of fable-5-premium, fable-5-premium-v2, fable-5.1-premium, CyberSec-Reasoning-Premium, and Kimi-K3-Premium.

Priorities: Quality > Ease of Access > Quantity

📊 Dataset Overview

Property	Value
Total Traces	13,044
Train Split	11,087 (85.0%)
Validation Split	978 (7.5%)
Test Split	979 (7.5%)
Average Quality	0.96 (0.5–1.0 band)
Teacher	Qwen3.8-27B (+ Qwen3.8-27B-GPTQ-4bit for Terminal-Bench 2.1)
Harness	terminus-2 (UkisAI) · OpenMP agent (Terminal-Bench 2.1)
Language	English only (CJK-filtered)
License	Apache-2.0 + MIT (per-source attribution below)
Created	2026-09-20

🔗 Sources

Source	Raw	Kept	License	Description
UkisAI (Qwen3.8-27B-multi-turn-agent-sft)	13,024	13,022	Apache-2.0	Qwen3.8-27B terminus-2 traces over OpenThoughts-Agent task base: 6.5k natural-language terminal tasks + 4.8k inferred-bug fixes, avg 13 turns, up to 64 turns
Lottolabs (terminal-bench-2.1-qwen3.8-27b-traces)	22	22	MIT	Long-horizon Terminal-Bench 2.1 agent trajectories (42+ turns, bash/read tool loops) — omp wire logs parsed to clean messages with thinking blocks and validated tool IDs
Excluded: open-thoughts/OpenThoughts-Agent-v1-SFT	—	—	Apache-2.0	All 15,209 rows are GLM-4.6 teacher (not Qwen3.8) — wrong teacher for this dataset
Excluded: yangw1234/swe-explore	—	—	—	Gated (manual)

🧹 Quality Pipeline

Deduplication — Cross-source SHA-256 hashing on normalized messages JSON
Wire-Log Parsing — OpenMP session logs converted to clean messages (thinking blocks preserved, tool calls re-serialized with IDs)
Language Filtering — CJK-ratio check for English-only consistency
PII Scrubbing — Local paths, API keys (sk-, hf_, AKIA), emails, IPs replaced with [REDACTED_*]
Content Filtering — Refusal patterns and short (<500 char) records removed
Quality Scoring — Length, structure, tool-loop depth composite (mean 0.96)

📈 Coverage

Natural-language terminal tasks · inferred-bug reproduction/fixes · long-horizon multi-step agent loops (bash, read, write tools) · regex/chess, torch pipeline parallelism, JS filtering and more Terminal-Bench 2.1 tasks

🎯 Usage

from datasets import load_dataset

ds = load_dataset("saidutta69/Qwen3.8-Agent-Premium", data_files="openai_chat/train.parquet", split="train")

With Axolotl

datasets:
  - path: saidutta69/Qwen3.8-Agent-Premium
    type: chat_template
    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. Companion datasets: fable-5-premium · fable-5.1-premium · CyberSec-Reasoning-Premium · Kimi-K3-Premium.

Downloads last month: 160