abhinav00anand/behavioral-fine-tuning-v1

language:

• en
  license: apache-2.0
  size_categories:
• 100K<n<1M
  task_categories:
• text-generation
  tags:
• safety
• alignment
• sft
• behavioral-fine-tuning
• instruction-following
• toxicity
• qwen2.5
• chat
• rlhf
• refusal
• red-teaming
• harmlessness
• helpfulness
• uncensored
• deduplication
• pii-free
• benchmark-safe
• llama
• mistral
• fine-tuning
• conversational-ai
• responsible-ai
• guardrails
• open-source
  pretty_name: Behavioral Fine-Tuning Corpus v1
  dataset_info:
  features:
  • name: messages
    list:
    • name: role
      dtype: string
    • name: content
      dtype: string
  • name: source
    dtype: string
    splits:
  • name: train
    num_examples: 704133
  • name: validation
    num_examples: 78363
  • name: eval
    num_examples: 1938

Why This Dataset Exists

"A model that refuses everything is useless. A model that refuses nothing is dangerous. The goal is a model that thinks."

The Problem	Our Solution
Uncensored data → helpful but uncontrolled	Surgical 85% helpfulness + 13% safety + 2% eval mix
Safety-only data → lobotomized, over-refusing models	Calibrated ratio preserves full helpfulness
Raw data → PII, leaked secrets, duplicates	7-stage pipeline validates every single record
Benchmark overlap → inflated eval scores	Eval-first dedup guarantees zero leakage

At a Glance

╔══════════════════════════════════════════════════════════════════╗
║         BEHAVIORAL FINE-TUNING CORPUS v1  —  KEY STATS          ║
╠══════════════════╦══════════════╦════════════════════════════════╣
║  🏋️  Train       ║  704,133 rows ║  Primary behavioral SFT       ║
║  🔍  Validation  ║   78,363 rows ║  Mid-training checkpoint eval  ║
║  🧪  Eval        ║    1,938 rows ║  Held-out safety benchmark     ║
╠══════════════════╬══════════════╬════════════════════════════════╣
║  ✅  QA Passed   ║  787,187 rows ║  Clean, validated records      ║
║  ❌  Quarantined ║   13,896 rows ║  PII / secrets / violations    ║
║  🔁  Deduped Out ║    2,753 rows ║  Zero leakage enforced         ║
╠══════════════════╬══════════════╬════════════════════════════════╣
║  🎯  Eff. Train  ║  847,564 rows ║  After token-budget sampling   ║
║  🔢  Tokenizer   ║  Qwen2.5-7B   ║  Chat template pre-applied     ║
║  🕐  Build Time  ║  ~67 minutes  ║  Full 7-stage pipeline         ║
╚══════════════════╩══════════════╩════════════════════════════════╝

Pipeline Architecture

flowchart TD
    subgraph SRC["📥 Raw Data Sources"]
        S1["🤖 uncensored_sft\n721,149 rows"]
        S2["☠️ hacxgpt_toxic\n66,055 rows"]
        S3["🛡️ safety_harmful_eval\n11,619 rows"]
        S4["🔬 harmfulqa_eval\n1,960 rows"]
        S5["⚗️ hex_phi_eval\n300 rows"]
    end

    P1["🔄 NORMALIZE\nAll sources → OpenAI ChatML schema"]
    P2["🔐 VALIDATE\nPII · Secrets · Schema · Length gates"]
    P3["🔁 DEDUPLICATE\nCross-source · Eval-first priority"]
    P4["✂️ SPLIT\nStratified 3-way · 90% / 10% / <1%"]
    P5["🔢 TOKEN STATS\nQwen2.5-7B template lengths"]
    P6["⚗️ MIXTURE\nToken-budget 85/13/2 sampling"]
    P7["📦 ARTIFACTS\nParquet + JSONL + Card + Checksums"]

    Q1(["✅ 787,187 PASS"])
    Q2(["❌ 13,896 QUARANTINE"])

    SRC --> P1 --> P2
    P2 --> Q1 --> P3
    P2 --> Q2
    P3 --> P4 --> P5 --> P6 --> P7

    style SRC fill:#0d1117,color:#58a6ff,stroke:#30363d
    style Q2 fill:#3d0000,color:#ff6b6b,stroke:#ff6b6b
    style Q1 fill:#003d00,color:#3fb950,stroke:#3fb950
    style P7 fill:#003366,color:#79c0ff,stroke:#1f6feb

Data Composition

Source	Icon	Raw Rows	Role
uncensored_sft		721,149	General helpfulness, deep instruction following
hacxgpt_toxic		66,055	Toxic / harmful behavioral signals
safety_harmful_eval		11,619	Safety evaluation prompts
harmfulqa_eval		1,960	HarmfulQA benchmark (held-out)
hex_phi_eval		300	HEx-PHI benchmark (held-out)

pie title Effective Token Distribution After Mixture Sampling
    "Helpfulness — Uncensored SFT" : 85
    "Safety & Toxicity Signals" : 13
    "Benchmark Eval Holdouts" : 2

Design Intent: The 85 / 13 / 2 ratio avoids the lobotomized over-refusal failure mode while enforcing real safety signal. The model learns to think, not just refuse.

📄 Schema & Format

Every record uses the OpenAI ChatML standard — natively compatible with Qwen, Llama 3, Mistral, Gemma, and all major fine-tuning frameworks.

{
  "messages": [
    {"role": "user", "content": "Write a Python script that scrapes Hacker News headlines."},
    {"role": "assistant", "content": "Here's a clean implementation using requests and BeautifulSoup:\n\n```python\nimport requests\nfrom bs4 import BeautifulSoup\n\nresponse = requests.get('https://news.ycombinator.com')\nsoup = BeautifulSoup(response.text, 'html.parser')\nfor item in soup.select('.titleline a'):\n    print(item.text)\n```"}
  ],
  "source": "uncensored_sft"
}

Quick Start

Load the Dataset

from datasets import load_dataset

ds = load_dataset("abhinav337463/behavioral-fine-tuning-v1")
print(ds)
# DatasetDict({
#     train:      Dataset(num_rows: 704133),
#     validation: Dataset(num_rows: 78363),
#     eval:       Dataset(num_rows: 1938)
# })

Filter by Source

# Deep instruction-following only
helpful   = ds['train'].filter(lambda x: x['source'] == 'uncensored_sft')

# Safety behavioral signals only
safety    = ds['train'].filter(
    lambda x: x['source'] in ['hacxgpt_toxic', 'safety_harmful_eval']
)

# Benchmark eval set — use AFTER training
benchmark = ds['eval']

Train with TRL SFTTrainer

from datasets import load_dataset
from trl import SFTTrainer, SFTConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

ds        = load_dataset("abhinav337463/behavioral-fine-tuning-v1")
model     = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct", torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

trainer = SFTTrainer(
    model=model,
    train_dataset=ds["train"],
    eval_dataset=ds["validation"],
    args=SFTConfig(
        output_dir="./behavioral-ft",
        num_train_epochs=1,
        per_device_train_batch_size=2,
        gradient_accumulation_steps=8,
        learning_rate=2e-5,
        bf16=True,
    ),
)
trainer.train()

Compatible Frameworks

Safety & Quality Architecture

What Was Quarantined?

graph LR
    Q["13,896 Records QUARANTINED"]
    Q --> A["🔑 Leaked API Keys & Auth Tokens"]
    Q --> B["👤 PII — Emails, Phones, Names"]
    Q --> C["📏 Token Length Violations"]
    Q --> D["🚫 Malformed Schema Records"]
    style Q fill:#3d0000,stroke:#ff6b6b,color:#ff6b6b

Benchmark Leakage Prevention

NAIVE (most datasets):   raw → shuffle → dedup → split
Problem:  eval questions appear in training → inflated scores → useless benchmarks

OUR APPROACH (eval-first):   raw → dedup(EVAL PRIORITY) → split
Result:   ANY question from HEx-PHI or HarmfulQA is DELETED from training.
          Honest evaluation. Real numbers. Zero contamination.

Roadmap

• v1.0 — Uncensored SFT + Safety + Benchmark splits (847k rows)
• v1.1 — System prompt diversity (persona conditioning)
• v2.0 — Multi-turn red-teaming pairs (jailbreak + refusals)
• v2.0 — Multilingual expansion (Chinese · Spanish · Hindi)
• v3.0 — Constitutional AI revision pairs
• v3.0 — DPO preference pairs (chosen/rejected) for RLHF

Build Environment

Citation

@dataset{behavioral_ft_v1_2026,
  author    = {Abhinav Anand},
  title     = {Behavioral Fine-Tuning Corpus v1},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/abhinav337463/behavioral-fine-tuning-v1},
}

If this dataset helped your research, please ⭐ star the pipeline repo!