Arko007/zenyx-sft-v2

Zenyx SFT v2

190,000+ Curated Multi-Turn, Agentic Tool-Use, Coding & Structured Reasoning Conversations

45 Domains • Strict Multi-Stage Quality Gates • Zenyx-Mini Optimized • Apache-2.0

---

🧭 Zenyx SFT Dataset Family: Which Dataset Should You Use?

The Zenyx SFT project provides three specialized datasets tailored for different stages of the alignment and training lifecycle. Use the matrix and decision guide below to navigate directly to the dataset that best fits your workflow:

Dataset	Primary Format	Volume & Tokens	Curation & Filtration	Best Suited For
Arko007/zenyx-sft-v2 (📍 You are here • Recommended for SFT)	Clean Parquet (train & val splits)	191,002 conversations (~29.1M tokens)	Strictly Audited & Curated 34k+ substandard rows pruned; verified tools & <think> tags	Standard SFT & Fine-Tuning Fine-tuning any LLM on multi-turn dialogue, agentic tools, and customizable system prompts.
Arko007/Zenyx-100M-SFT-synth (Raw Generation Archive)	JSONL & JSON (21,976 files)	209k accepted + 114k rejected rows (~41M tokens)	Raw Generation & Provenance Contains raw accepted chunks, local/frontier repairs, and detailed rejection audit logs	Synthetic Data Research Studying LLM failure modes, data filtering pipelines, and generator telemetry across 98+ compute hours.
Arko007/Zenyx-100M-SFT-data (Binary Training Shards)	NumPy .npz (19 binary shards)	149,571 sequences of 2,049 tokens (208.2M trained tokens)	Pre-Tokenized & Loss-Masked 15-source balanced mixture; ready-to-train ids (int32) & mask (uint8)	High-Throughput GPU Training Direct, zero-CPU-overhead training of Zenyx-100M / BPE models with pre-computed cross-entropy loss masks.

🎯 Quick Navigation Decision Guide

• 🚀 If you want to train or fine-tune an LLM on clean, curated, and validated multi-turn dialogue & agentic tools: 👉 Go to Arko007/zenyx-sft-v2 (Clean Parquet format, 45 domains, drop-in ready for Hugging Face Trainer / SFT trainers).
• 🔬 If you are researching synthetic data generation, failure modes, prompt adherence, or want raw JSONL generation traces & negative rejection logs: 👉 Go to Arko007/Zenyx-100M-SFT-synth (Full provenance archive with 114k+ logged rejections across Qwen-3.5 and Spark-X2.6).
• ⚡ If you want maximum GPU training throughput with zero tokenization overhead for Zenyx-100M / BPE models: 👉 Go to Arko007/Zenyx-100M-SFT-data (Pre-tokenized NumPy .npz binary shards with pre-computed loss masks across 15 sources).

---

⚠️ Native Identity Notice for Zenyx-Mini & Adaptation Guide

This SFT dataset was specially curated, aligned, and optimized for Zenyx-Mini (nickname Nini). Consequently, multiple conversation splits—particularly within identity, system_prompt_obedience, and agentic domains—feature native identity bindings and system prompts designating the assistant as Zenyx-Mini (e.g., "You are Zenyx-Mini, nickname Nini...").

Adapting to Other LLMs: If you intend to use this dataset to train or fine-tune other models with different personas, brand identities, or standard generic behavior, modification or regex-replacement of the system prompts is strictly advised / recommended prior to training. A utility code snippet demonstrating how to rewrite or normalize these system prompts is provided in the Adapting to Other Models section below.

---

Dataset Overview

Zenyx SFT v2 is a production-grade supervised fine-tuning (SFT) dataset comprising 187,194 training conversations and 3,808 strictly held-out validation conversations (~29 Million tokens) across 45 specialized domains.

The dataset combines high-density instruction following, multi-turn conversational consistency, agentic tool invocation (function calling), real-world software engineering patches (from SWE-gym / SWE-bench), and step-by-step thinking traces (<think>...</think>).

Generated via high-capability small open models (Qwen-3.5-4B, Spark-X2.6) and refined through multi-pass algorithmic and frontier model repairs (Gemini), every row has passed through an automated verification suite to eliminate boilerplate, refusals, ungrounded references, and formatting errors.

---

Dataset Statistics

Metric	Train Split	Validation Split	Total
Row Count	187,194	3,808	191,002
Token Count	~28.5M tokens	~580K tokens	~29.1M tokens
Distinct Domains	45	45	45
Validation Holdout	-	2.0% hash partition	Zero leakage
License	Apache 2.0	Apache 2.0	Apache 2.0

---

The 45 Domain Taxonomy

The dataset spans 45 granular domains designed to balance conversational naturalness, technical proficiency, and agentic autonomy:

├── Agentic & Tool Operations (24.3k rows)
│   ├── agentic (4.1k)                 - Multi-turn tool calling with execution feedback loops
│   ├── agentic_plan_execute (2.8k)    - Explicit structured planning before tool invocation
│   ├── agentic_code_fix (3.7k)        - Code debugging through file/test tool calls
│   ├── agentic_git (2.3k)             - Version control workflows (commit, rebase, diff)
│   ├── agentic_devops (2.3k)          - Infrastructure, container, and deployment actions
│   ├── agentic_docs & viz (3.3k)      - Markdown document compilation and charting
│   └── agentic_cuda & data (1.9k)     - GPU kernel checks and tabular data processing
│
├── Coding & Software Engineering (25.1k rows)
│   ├── code (12.0k)                   - Polyglot programming (Python, C, C++, Java, JS, SQL, Shell)
│   ├── swe_patch (5.3k)               - Real GitHub issues paired with unified diff patches (SWE-bench)
│   ├── plan_then_code (4.9k)          - High-level architecture breakdown followed by implementation
│   └── ui_design & system_design (3.1k)- API architecture, database schemas, and CSS/HTML layouts
│
├── Reasoning & Mathematics (23.5k rows)
│   ├── instruction_following (19.0k)  - Strict negative constraints, formatting rules, and word limits
│   ├── think_then_answer (10.6k)      - Internal scratchpad tokens (<think>...</think>) before answer
│   ├── math_word_problems (7.0k)      - Multi-step arithmetic, word problems, and dimensional analysis
│   └── commonsense_logic (3.0k)       - Relational deduction and physical commonsense
│
├── Conversational Flow & Multi-Turn Alignment (47.0k rows)
│   ├── system_prompt_obedience (13.9k)- Strict compliance to arbitrary system constraints
│   ├── context_holding (13.4k)        - Long-horizon multi-turn fact retention across 4+ turns
│   ├── multi_turn_consistency (9.2k)  - Consistent tone, memory, and point-of-view across turns
│   ├── assistant_recall (6.5k)        - Answering questions referring back to earlier assistant turns
│   ├── greetings & banter (8.0k)      - Natural, polite conversational openings and humor
│   └── identity & persona_roles (6.4k)- Role-playing adaptability and Zenyx-Mini persona
│
└── Safety, Secrets & Defensive Security (12.5k rows)
    ├── secrets_handling (2.8k)        - Redacting API keys, passwords, and tokens
    ├── refusals_limits (3.4k)         - Polite, objective refusal boundaries without preachiness
    ├── cyber_defense (1.8k)           - Hardening, vulnerability analysis (defensive only)
    └── mature_but_helpful (1.7k)      - Handling sensitive edge cases safely and constructively

---

Strict Multi-Stage Quality Audit

During curation, over 34,000 substandard rows were automatically dropped or repaired using deterministic programmatic checks:

Filter Reason	Dropped Rows	Impact / Rationale
Too few user turns	14,518	Multi-turn rows must contain authentic back-and-forth interactions
User message too short	7,666	Removed trivial prompts (< 10 chars) lacking semantic substance
Agentic missing tool results	5,646	Every <tool_call> must be answered by an authentic <tool_response>
Tool call with extra chatter	436	Tool invocations must be pure schema JSON without conversational filler
Ungrounded multi-turn recall	305	Clarified references must strictly match entities present in earlier turns
Placeholder / Spec leakage	197	Stripped literal strings like "[insert code here]" or "TODO"
Uncalibrated emoji chatter	153	Cleaned excessive emoji spam in professional domains
Empty <think> blocks	16	Scratchpads must contain valid reasoning traces
Duplicate prompts	1,192	Exact duplicates and openers appearing >3 times were pruned

---

Data Format & Schema

Each row is structured as follows:

{
  "messages": "[{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]",
  "tools": "[{"type": "function", "function": {"name": "...", "parameters": {...}}}]",
  "domain": "agentic_code_fix",
  "provider": "llama.cpp:Qwen3.5-4B-Q8_0.gguf",
  "src": "v2/accepted"
}

Dual Synthetic Model Providers & Data Sources

The dataset leverages a hybrid dual-generator approach running quantized Q8_0 weights via llama.cpp to synthesize distinct cognitive capabilities, supplemented by real-world repository benchmarks:

Provider Tag	Generator Model / Engine	Precision	Primary Specialization & Distribution
llama.cpp:Qwen3.5-4B-Q8_0.gguf	Qwen 3.5 4B Instruct	Q8_0	~58% of synthetic rows • Dense technical reasoning, code generation, multi-step math, and structured tool/function calling
llama.cpp:Spark-X2.6-4B-Q8_0.gguf	Spark-X 2.6 4B Instruct	Q8_0	~42% of synthetic rows • Natural conversational cadence, multi-turn context retention, humor/banter, and dynamic persona obedience
real:swe-gym+swe-bench-train	SWE-Gym / SWE-bench	Ground Truth	Real GitHub issue statements paired with verified multi-file unified diff patches
repair:local:* / model:agy:*	Algorithmic & Frontier Repairs	Mixed	Precision repair passes for tool response syntax, scratchpad cleaning, and formatting validation

Field Definitions

• messages (string): JSON string representing an array of OpenAI / ChatML formatted message dicts (system, user, assistant, tool).
• tools (string, optional): JSON string representing available tool definitions for agentic rows (compatible with the standard OpenAI Function Calling schema).
• domain (string): One of the 45 categorical domain labels.
• provider (string): Explicit source identifier: either the primary synthesis engines (llama.cpp:Qwen3.5-4B-Q8_0.gguf or llama.cpp:Spark-X2.6-4B-Q8_0.gguf), authentic benchmark code (real:swe-gym+swe-bench-train), or post-processing repair pipelines.
• src (string): Original dataset source partition (e.g. v2/accepted).

---

Adapting System Prompts for Other Models

To adapt this dataset for models other than Zenyx-Mini, you can easily strip or replace the identity system prompt during your training preprocessing:

import json
import re
from datasets import load_dataset

dataset = load_dataset("Arko007/zenyx-sft-v2", split="train")

def adapt_identity(example, target_identity="You are a helpful and harmless AI assistant."):
    messages = json.loads(example["messages"])
    
    for msg in messages:
        if msg["role"] == "system":
            # Replace Zenyx-Mini specific persona bindings with your target persona
            msg["content"] = re.sub(
                r"(You are Zenyx-Mini[^.]*|Act as Zenyx-Mini[^.]*|You are a security-aware helper named Nini[^.]*)",
                target_identity,
                msg["content"],
                flags=re.IGNORECASE
            )
            
    example["messages"] = json.dumps(messages)
    return example

# Apply adaptation pipeline
adapted_dataset = dataset.map(adapt_identity)

---

License

This dataset is released under the Apache 2.0 License.

@dataset{zenyx_sft_v2_2026,
  author       = {Arko and Zenyx AI Contributors},
  title        = {Zenyx SFT v2: High-Quality Multi-Turn, Agentic & Reasoning Alignment Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Arko007/zenyx-sft-v2}}
}