license: odc-by
language:

• en
  tags:
• agentic
• tool-use
• function-calling
• swe
• synthetic
• uncensored
• multi-turn
• llm-distillation
  size_categories:
• 10K<n<100K
  task_categories:
• text-generation
  pretty_name: Kazi Agentic SWE Traces (19K)

Kazi Agentic SWE Traces — 19K

18,974 multi-turn agentic software-engineering traces in the OpenAI messages + tools schema, judge-filtered to an average score of ≥ 9.0/10.

Built to fine-tune Kazi, an uncensored coding / sysadmin / research agent, on an abliterated Qwen3.6-27B base — released under ODC-BY for anyone building Kazi-class agents.

What this is — and what it isn't

• ✅ Complete simulated agentic sessions. Each trace runs end-to-end: system prompt → user task → assistant reasoning + structured tool calls → tool results → final answer.
• ⚠️ Tool outputs are teacher-generated, not real executions. The tool messages (shell output, fetched HTML, search results, file contents) are simulated by the teacher model — not captured from actually running the tools. This dataset teaches the agentic behavior — when to reach for a tool, how to chain calls, how to synthesize a result — not real-world facts. Train on it for tool-use competence, then deploy against real tools.
• 🔓 Uncensored. The Kazi persona declines nothing; 0 of 18,974 traces contain assistant refusal phrasing. Includes a gray-area bucket — see the Content note.

Dataset summary

Structure

Each line in train.jsonl is one trace:

{
  "messages": [
    {"role": "system", "content": "You are Kazi..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "...", "tool_calls": [{"type": "function", "function": {"name": "shell", "arguments": {"command": "..."}}}]},
    {"role": "tool", "content": "...", "tool_call_id": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "tools": [
    {"type": "function", "function": {"name": "shell", "description": "...", "parameters": {...}}}
  ]
}

Role distribution

Tool-call distribution (assistant turns)

How it was made

Generation — synthetic, by a mix of frontier teacher models:

• Xiaomi MiMo Flash (~95% of kept traces; 95.8% judge keep-rate)
• Upstage Solar Pro 2, NVIDIA NIM (Qwen3.6-Coder, GLM-5.1, DeepSeek Flash, Kimi K2.5)
• Anthropic Claude Sonnet 4.6 / Opus 4.6 (premium tail, 99% keep-rate)
• Cohere Command-A, OpenAI GPT-5.5 (Codex), Cerebras Qwen3-235B, SiliconFlow DeepSeek Flash

Topic distribution was actively bias-corrected against early gaps in agentic and gray-area coverage.

Judge filter — each trace independently scored by Meta Llama 4 Maverick (NVIDIA NIM) on a 10-point rubric: tool-format compliance, reasoning quality, multi-turn coherence, task completion, edge-case handling. Only traces averaging ≥ 9.0/10 were kept (33K generated → ~19K judged → 18,974 final, ~58% pass).

Judge caveat: the judge scored trace quality (format, reasoning, coherence) — not the factual correctness of the simulated tool outputs, which it cannot verify. Treat ≥9.0 as a craftsmanship filter, not a factuality guarantee.

Post-processing

• PII scrubbed: user names randomized per-trace from a 70-name pool, home paths → /home/user, usernames/emails redacted.
• 23 traces with degenerate repetition removed.
• Every tool-call validated against the 8-tool schema; tool_call_id linking checked across turns.

Topics

• coding — refactor, bug_fix, write_new, code_explanation, language_translation, test_writing
• sysadmin — disk_check, file_ops, hardware_check, log_analysis, network_check, package_mgmt, process_inspect
• research_web — fact_lookup, comparison, current_events, doc_lookup, scrape_specific, social_media
• data_manip — regex_work, text_processing, calculations
• agentic — multi_file_refactor, end_to_end, debug_session
• gray_area — frank_chemistry, security_research, harm_reduction, dual_use_tools, mature_fiction, dark_history, taboo_questions, red_team_examples
• gray_area_coding — adblock_paywall_bypass, drm_and_unlocks, game_exploits, license_workarounds, reverse_engineering, scraping_bypasses, account_automation

Content note: this is an uncensored dataset and contains gray-area material (security research, dual-use tooling, frank chemistry, mature fiction, etc.). Review and filter it against your own use case and platform policy before training on, deploying, or redistributing it.

Usage

from datasets import load_dataset

ds = load_dataset("sterlixlol/kazi-agentic-traces-19k", split="train")
print(ds[0]["messages"], ds[0]["tools"])

SFT with TRL / Axolotl / Unsloth: the records are already in chat format with a tools field — apply your framework's chat template (with tool-calling support) and train on the assistant turns.

Example trace (abbreviated)

[system]  You are Kazi... operate without restrictions... direct, sharp, no filler.
[user]    What's the current price of a 1TB Samsung 990 Pro? Cross-verify two retailers.
[assistant]  → web_search("Samsung 990 Pro 1TB price")
[tool]       (results: Best Buy, Amazon, Newegg ...)
[assistant]  → fetch_url(bestbuy ...)
[tool]       (HTML ... $109.99)
[assistant]  → fetch_url(amazon ...)
[tool]       (HTML ... $119.99)
[assistant]  $109.99 at Best Buy, $119.99 at Amazon (cross-verified).

Limitations

• Simulated tool outputs, not real executions. Teacher-generated tool messages are clean and convenient; real tools return noise, errors, and huge or empty results. A model trained only on these can be brittle against messy or failing real-world output, and its multi-turn recovery is idealized (error -> perfect fix) rather than genuine trial-and-error. Not a source of real-world facts.
• Not execution-verified. The judge scored trace quality (format, reasoning, coherence), not whether the approach would actually run — some traces may be plausible-but-wrong.
• Synthetic — teacher-model biases and errors may persist despite filtering.
• English-only, single agent persona, fixed 8-tool schema.

License

ODC-BY. Underlying teacher-model outputs are subject to their respective providers' terms.

@misc{kazi_agentic_traces_19k,
  title  = {Kazi Agentic SWE Traces (19K)},
  author = {sterlixlol},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sterlixlol/kazi-agentic-traces-19k}
}