Datasets:

saidutta69
/

RaceBench-v1.1

messages large listlengths 2 246	source_group large_stringclasses 2 values	source large_stringclasses 15 values	category large_stringclasses 59 values	quality_score float64 60 100
[ { "role": "user", "content": "请分析这段安全相关的文本内容" }, { "role": "assistant", "content": "安全文本分析：\n\n# Baby RSA (crypto, 74p, 137 solved) In the challenge we get [code](rsa.sage), the [public key](pubkey.py) and the [result](flag.enc). Encryption is rather simple: ```python R.<a> = GF(22049) def encry...	distillation	high_quality_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "What techniques prevent subdomain takeover through implementation of DNS firewalls and anomaly detection systems?" }, { "role": "assistant", "content": "Subdomain takeover prevention requires implementing DNS firewalls and anomaly detection systems that monitor and ...	distillation	fenrir_v2_1	cybersecurity	60
[ { "role": "user", "content": "Generate a hierarchical list ranking from most utilized to least utilized containing the common techniques and methodologies utilized by contemporary installation artists in order to produce immersive conceptual art pieces that suspend viewers' perceptual assumptions, indicatin...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "You are tasked with implementing a file management class in PHP. Your task is to create a method `moveFile` that moves a file from its current location to a new specified location. The method should perform the following checks and operations:\n\n1. Check if the source file exi...	distillation	kimi_coding	coding	60
[ { "role": "user", "content": "title:Poster: On the Application of NLP to Discover Relationships between" }, { "role": "assistant", "content": "title:Poster: On the Application of NLP to Discover Relationships between\nMalicious Network Entities\nauthor:Giuseppe Siracusano and\nMartino Trevisan a...	distillation	high_quality_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "Calculate the total amount of precipitation over a three month period from rain gauge measurements across 10 different locations using a weighted average that takes into account the surface area represented by each gauge, as well as seasonal precipitation norms for the region. ...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "<!DOCTYPE html>" }, { "role": "assistant", "content": "<!DOCTYPE html>\n<html lang=\"en\">\n<head>\n <meta charset=\"utf-8\">\n <title>JSDoc: Source: browser/cookie.js</title>\n\n <script src=\"scripts/prettify/prettify.js\"> </script>\n <script src=\"sc...	distillation	high_quality_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "How does cyber kill chain mapping assist in anticipating subsequent stages after an initial Threat Informed Defense foothold?" }, { "role": "assistant", "content": "Cyber kill chain mapping provides structured visibility into adversary progression from initial compr...	distillation	fenrir_v2_1	cybersecurity	60
[ { "role": "user", "content": "title:\"Mix-in-Place\" anonymous networking using secure function evaluation" }, { "role": "assistant", "content": "title:\"Mix-in-Place\" anonymous networking using secure function evaluation\nauthor:Nilesh Nipane and\nItalo Dacosta and\nPatrick Traynor\n“Mix-In-Pl...	distillation	high_quality_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "Explain in detail the concept of reasoning and what differentiates deductive reasoning from inductive reasoning. Then evaluate how each method of reasoning impacts the reliability of conclusions by formulating examples of logistic functions which model the reliability r of conc...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "Provide an example code snippet in any programming language that creates and prints a random alphanumeric string of 50 characters in length. Be sure to include relevant comments explaining the code. \n\nUnless the language has restrictions on the maximum length of strings, in w...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "How to develop microarchitectural side-channel monitoring for zero-day detection?" }, { "role": "assistant", "content": "Developing microarchitectural side-channel monitoring for zero-day detection requires sophisticated approaches that leverage hardware-level obser...	distillation	fenrir_v2_1	cybersecurity	60
[ { "role": "user", "content": "Design a highly optimized, saga pattern for application block 5208 in a martech domain using Python. The system must integrate with Kafka services, handle livelock scenarios gracefully, and maintain lock-free constraints. Include synchronization primitives for Mod153 across App...	distillation	gemini35	distilled	60
[ { "role": "user", "content": "Impressionable Partridge 238500" }, { "role": "assistant", "content": "{\"id\": \"4a409ef0-9fb5-11ee-b57e-09f55271957e\", \"partition\": \"43708a90-9fb5-11ee-b57e-09f55271957e\", \"nodes\": {\"b3bc1064-0624-11ee-b173-c9921719c2f8\": {\"id\": \"HOST-2823\", \"partiti...	distillation	precinct6_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "In what systematic way would you break down an unfamiliar, complex concept into simpler concepts and use analogies, examples, and thought experiments to help a friend build an intuitive understanding of the idea? Provide specific strategies you could employ and the types of cla...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "How can you use deception technologies to detect native tool lateral movement attempts?" }, { "role": "assistant", "content": "Deception technologies effectively detect native tool lateral movement by creating realistic but monitored environments that mimic legitima...	distillation	fenrir_v2_1	cybersecurity	60
[ { "role": "user", "content": "Icy Ladybug 237261" }, { "role": "assistant", "content": "{\"id\": \"4810d7b0-7726-11ee-af8e-272b0e07eabf\", \"partition\": \"3d9399d0-7726-11ee-af8e-272b0e07eabf\", \"nodes\": {\"da6b9554-062e-11ee-9e70-7d7711d897d8\": {\"id\": \"HOST-2711\", \"partition\": \"9ba1a...	distillation	precinct6_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "Implement an S2PL manager in Python that uses 'Radix Trees' to store lock keys. This is intended for a system with millions of fine-grained locks. Constraints: Optimize the tree for prefix-based range locking. Implement a custom memory allocator in Python (using a large bytearr...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "How can security teams implement nested virtualization to transparently analyze VM-aware malware samples?" }, { "role": "assistant", "content": "# Implementing Nested Virtualization for VM-Aware Malware Analysis\\n\\n## Architecture and Implementation Strategy\\n\\n...	distillation	fenrir_v2_1	cybersecurity	60
[ { "role": "user", "content": "Write a cohesive 600 - 800 word imaginative story describing in precise detail how you master the skill of origami within one week while working a full time job that requires focus under tight deadlines; then proceeding to invent and demonstrate at least five complex origam...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "Generate six new random digital circuit designs, each containing between 5 to 15 electronic components such as resistors, capacitors, transistors, integrated circuits, and wires. For each circuit, describe what function the design performs and explain how changing one or more o...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "Prickly Bug 231053" }, { "role": "assistant", "content": "{\"id\": \"72e45750-abde-11ed-bb52-0528ab1bf559\", \"partition\": \"6d2fec20-abde-11ed-bb52-0528ab1bf559\", \"nodes\": {\"f4975e20-a097-11ed-b6df-0fd1db319457\": {\"id\": \"f4975e20-a097-11ed-b6df-0fd1db31945...	distillation	precinct6_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "Design a highly optimized, strategy pattern for application block 4920 in a healthcare domain using Go. The system must integrate with Kafka services, handle livelock scenarios gracefully, and maintain constant memory constraints. Include synchronization primitives for Mod54 ac...	distillation	gemini35	distilled	60
[ { "role": "user", "content": "Design a complete, fact-based training recipe to turn Claude-3.5-Sonnet into a stronger Recursive Seed AI focused on \"reduce inference latency by 3x\" with measurable gains on DS-1000.\n\nRequirements:\n- Use only published, reproducible techniques.\n- Specify exact hyperpara...	distillation	gemini_pro32	distilled	60
[ { "role": "user", "content": "Calculate the amount of potential kinetic energy in Joules for a 80kg sprinter 2 seconds into a sprint from rest assuming frictionless conditions and an acceleration of 3m/s2. Express your final answer with two significant figures. Also take into account any situation where th...	distillation	fable5_agentic_sft	coding	60
[ { "role": "user", "content": "title:Towards asymmetric searchable encryption with message recovery and" }, { "role": "assistant", "content": "title:Towards asymmetric searchable encryption with message recovery and\nflexible search authorization\nauthor:Qiang Tang and\nXiaofeng Chen\nTowards Asy...	distillation	high_quality_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "Extensive Tiger 243369" }, { "role": "assistant", "content": "{\"id\": \"ebbc7870-168a-11ef-a93b-c5a8e87bdfec\", \"partition\": \"ebb8a7e0-168a-11ef-a93b-c5a8e87bdfec\", \"nodes\": {\"24d1ece4-0628-11ee-bea9-5f52a09eb73a\": {\"id\": \"HOST-2821\", \"partition\": \"9...	distillation	precinct6_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "How can machine learning improve dynamic analysis accuracy and reduce false positives?" }, { "role": "assistant", "content": "Machine learning significantly enhances dynamic analysis accuracy and reduces false positives through sophisticated pattern recognition, beh...	distillation	fenrir_v2_1	cybersecurity	60
[ { "role": "user", "content": "## SecCoding (Misc, 100p)" }, { "role": "assistant", "content": "## SecCoding (Misc, 100p)\n> You should fix vulnerabilities of the given source code, WITHOUT changing its normal behaviour. Link \n###ENG\n[PL](#pl-version)\nWe are given following code, and are tas...	distillation	high_quality_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "# Caesar's favourite song (Misc/Crypto/Stegano)" }, { "role": "assistant", "content": "# Caesar's favourite song (Misc/Crypto/Stegano)\nIn the task we get a [song](killthegauls.mp3).\nWe also know that the flag format is `DCTFXXXDCTF`.\nWe start off by translating t...	distillation	high_quality_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "Glorious Cheetah 242014" }, { "role": "assistant", "content": "{\"id\": \"651ae6b0-ea27-11ee-9c4f-25cb7d94c0ce\", \"partition\": \"5f31fe00-ea27-11ee-9c4f-25cb7d94c0ce\", \"nodes\": {\"baeaf120-0625-11ee-bea9-5f52a09eb73a\": {\"id\": \"HOST-7117\", \"partition\": \"...	distillation	precinct6_cybersecurity	cybersecurity	60
[ { "role": "user", "content": "Determine the set of all real values for the parameter $a$ such that the function $f(x) = ax - \\ln(1 + x^2)$ is monotonically increasing on the set of real numbers." }, { "role": "assistant", "content": "To find the values of $a$ for which $f(x) = ax - \\ln(1 + x^2...	distillation	fable5_agentic_sft	coding	60
[ { "role": "system", "content": "You are an autonomous coding agent running on the user's machine.\n\nYou work in the current directory of a real repository. Use your tools to read, search, create, and edit files and to run shell commands (tests, builds, type-checkers). Some tools are deferred: only their na...	fable-5-premium	crownelius	agent	91.7

End of preview. Expand in Data Studio

RaceBench v1.1

A curated SFT dataset blending agentic tool-use traces with high-quality distillation data — purpose-built for making small models (0.5B-3B) capable enough to replace API-based frontier models in edge deployments.

Quality over quantity. Every row passed a quality threshold of >=60/100. v1.1 fixes the v1.0 agent dilution bug and upgrades to premium traces.

Dataset Composition

Blended from two source datasets:

saidutta69/fable-5-premium — agent traces with tool calls (>=0.8 filtered, premium)
saidutta69/GPT-5.5-...-Distillation-Cleaned — filtered to quality >= 60, categories: coding + cybersecurity + distilled

Source	Rows	Percentage
Fable-5 Premium Agent Traces (>=0.8, 4x upsampled)	23,280	6.2%
Distillation (coding + cybersecurity + distilled, quality >=60)	353,456	93.8%
Total	376,736	100%

v1.0 vs v1.1: v1.0 README claimed 25,460 agent (7.1%) but shipped 6,365 low-scored rows (1.77%, mean 0.87). v1.1 ships 5,820 unique premium rows (mean 89.7/100, min 80) correctly upsampled 4x to 23,280 — 3.65x more agent signal, all high-quality. Distillation unchanged (353,456).

Source 1: Fable-5 Premium Agent Traces

Multi-turn Claude Fable-5 Premium agent traces with tool calls, reasoning chains, and structured function-calling patterns. Filtered to overall >=0.8 (5,820 of 6,365, 91.4%) then upsampled 4x to preserve agentic signal. Mean quality 89.7, up from 0.87 (unfiltered).

Source 2: Distillation (Quality >= 60)

Filtered from the GPT-5.5 / Gemini / Grok / Claude Distillation Dataset, keeping only:

Coding — 135K rows with quality >= 60 (2.1% of 6.3M)
Cybersecurity — 167K rows with quality >= 60 (42% of 396K)
Distilled — 61K rows with quality >= 60 (93% of 65K)

Why not add instruction/science? Prime v2 instruction 52K only 2,008 >=60 (3.8%), science 74 >=60 (2.3%), humanities 0 >=60 — adding would be noise. Distilled already high-quality.

Quality Distribution

Distillation all ==60, premium 80-100 (mean 89.7). Overall mean 60.8 (v1 58.95). No rows <80 in agent split.

Why RaceBench for Edge Models?

Small models (0.5B-3B) need high-signal, low-noise data to punch above their weight. RaceBench:

Tool-use signal — Agent traces teach function calling and multi-step reasoning patterns that small models struggle to learn from plain instruction data
Quality floor — No row below quality 60/100 (distill) or 80/100 (premium agent); poor-quality rows are excluded
Domain focus — Coding + security + distilled is the sweet spot for practical edge deployment
Deduped — Cross-dataset SHA-256 dedup removed 29K redundant rows (v1), v1.1 shuffled seed 42, zstd compressed

Data Fields

Column	Type	Description
`messages`	`List[Struct{role, content}]`	Chat-format conversation (SFT-ready)
`source_group`	`str`	`"fable-5-premium"` or `"distillation"`
`source`	`str`	Original source split name
`category`	`str`	Content category (`agent`, `coding`, `cybersecurity`, `distilled`)
`quality_score`	`float64`	Quality score (0-100, distill ==60, premium 80-100)

Usage

from datasets import load_dataset

ds = load_dataset("saidutta69/RaceBench-v1.1", split="train")
print(f"Examples: {len(ds)}")  # 376736
print(ds[0]["messages"])

# Axolotl config
datasets:
  - path: saidutta69/RaceBench-v1.1
    type: chat_template
    split: train

# Filter by category
coding = ds.filter(lambda x: x["category"] == "coding")
agent = ds.filter(lambda x: x["category"] == "agent")
print(f"Agent rows: {len(agent)}")  # 23280

RaceBench-MiniCPM5 — the RaceBench reference model

RaceBench v1 has its first public reference checkpoint: saidutta69/RaceBench-MiniCPM5 — a full-parameter fine-tune of openbmb/MiniCPM5-1B trained on v1, released with fp16 weights and all GGUF quantizations (Q2_K to F16).

Head-to-head vs base, evaluated with lm-evaluation-harness v0.4.12 under identical settings (fixed seed, 2x Tesla T4). z = delta / pooled standard error; |z| > 2 is significant at ~95% confidence:

Task	RaceBench-MiniCPM5	MiniCPM5-1B (base)	Delta	z
BBH (zero-shot, 27 subtasks)	0.3397 ± 0.0056	0.3123 ± 0.0055	+2.74	+3.5
GSM8K	0.3108 ± 0.0127	0.4102 ± 0.0135	-9.93	-5.3
MMLU (subset)	0.5310 ± 0.0060	0.5497 ± 0.0059	-1.87	-2.2
Minerva MATH-500	0.2140 ± 0.0184	0.2620 ± 0.0197	-4.80	-1.8
Winogrande	0.5391 ± 0.0140	0.5620 ± 0.0139	-2.29	-1.2
HellaSwag	0.4820 ± 0.0050	0.4881 ± 0.0050	-0.61	-0.9
TruthfulQA MC2	0.4710 ± 0.0149	0.4597 ± 0.0149	+1.13	+0.5
GPQA Main (n-shot)	0.2589 ± 0.0207	0.2723 ± 0.0211	-1.34	-0.5
ARC-Challenge	0.3746 ± 0.0141	0.3831 ± 0.0142	-0.85	-0.4

What it shows: RaceBench transfers real multi-step reasoning to a 1B model (BBH +2.7, z=3.5 - the only significant gain). The price is a real math regression (GSM8K -9.9, z=-5.3) and a small knowledge loss (MMLU -1.9). All other deltas are within noise. v1.1 keeps the same distillation but with 3.65x more premium agent signal — expect BBH gain to hold, GSM8K regression not to worsen (premium is more reasoning-dense). Benchmark for v1.1 pending (same harness, same seed).

Training v1.1: openbmb/MiniCPM5-1B, full-parameter, fp16, sequence_len 4096, 1 epoch, lr 2e-05 cosine, same as v1. See /tmp/RaceBench-v1.1-finetune-README.md for GCP launch.

Citation

@misc{racebench,
  author = {Sai Dutta},
  title = {RaceBench -- Edge-Focused SFT Blend for Small Models},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/saidutta69/RaceBench}}
}

@misc{racebench-v1.1,
  author = {Sai Dutta},
  title = {RaceBench v1.1 -- Quality-Expanded (Premium Agent)},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/saidutta69/RaceBench-v1.1}}
}

Downloads last month: -