exnivo/Lume-Instruct-SFT

Lume Instruct SFT

A 160k+ row high-quality English SFT dataset for fine-tuning small instruction-following language models.

Lume Instruct SFT is a multi-domain supervised fine-tuning dataset engineered for small language models (around 100M–3B parameters).

The dataset blends cleaned synthetic instruction data with curated subsets from HuggingFaceTB/smoltalk (everyday conversations, strict constraint following, text rewriting, summarization, and system-prompted dialogue). All hallucinated RAG source citations ("The source says...", "According to the passage...") have been strictly filtered out.

Quick Start

Load the dataset directly with Hugging Face Datasets:

from datasets import load_dataset

ds = load_dataset("exnivo/Lume-Instruct-SFT", split="train")

print(ds)
print(ds[0])

At a Glance

Item	Details
Dataset Name	exnivo/Lume-Instruct-SFT
Total Rows	160,617
Language	English
Format	Chat messages with user, assistant, and system roles
License	Apache 2.0
Best For	Small LLMs around 100M–3B parameters
Primary Use	Instruction Fine-Tuning (SFT) & Conversational Alignment
Related Model	exnivo/Lume

Key Features & Improvements over V1

1. Zero Unprompted RAG Hallucinations:
   • Removed 58,000+ synthetic RAG rows where the assistant referred to imaginary passages/studies ("According to the source...") without context provided in the prompt.
2. Integrated HuggingFaceTB/smoltalk:
   • Added everyday chit-chat, greetings, constraint-following, text rewriting, and summarization datasets to give small models natural, polite, and well-structured conversational behavior.
3. Optimized for Small LLMs (100M–3B):
   • Short, direct assistant completions, clean formatting, and realistic token lengths.

---

Real Dataset Statistics

Category Breakdown

Category	Rows	Percentage	Description
math_reasoning	37,611	23.42%	Python-verified math & basic arithmetic reasoning
clean_conversation	34,257	21.33%	Direct multi-turn assistant dialogues & Q&A
messy_idea_to_plan	29,978	18.66%	Turning unstructured user ideas into actionable plans
honesty_uncertainty	14,956	9.31%	Admitting unknown facts and polite refusals
simplify_explain	11,562	7.20%	Explaining complex topics in simple terms
simple_coding	9,993	6.22%	Python/Lua/JS code generation and debugging
system_smalltalk	5,000	3.11%	System-prompted assistant dialogues (smoltalk)
constraints_following	5,000	3.11%	Strict formatting & rule compliance (smoltalk)
text_rewriting	5,000	3.11%	Text editing, tone adjustment, & proofreading (smoltalk)
text_summarization	5,000	3.11%	Document & article summarization (smoltalk)
everyday_smalltalk	2,260	1.41%	Casual chit-chat & greetings (smoltalk)
Total	160,617	100.00%

Turn Type Breakdown

Turn Type	Rows	Percentage
single_turn	101,481	63.18%
multi_turn	59,136	36.82%

Source Composition

Source	Rows	Percentage
lume_cleaned_sft	138,357	86.14%
smoltalk_systemchats	5,000	3.11%
smoltalk_constraints	5,000	3.11%
smoltalk_rewrite	5,000	3.11%
smoltalk_summarize	5,000	3.11%
smoltalk_everyday	2,260	1.41%

---

Dataset Structure

Each row is a standard chat conversation dictionary.

{
  "category": "math_reasoning",
  "messages": [
    {
      "role": "user",
      "content": "What is 40% of 200?"
    },
    {
      "role": "assistant",
      "content": "40% of 200 is 80. Multiply 200 by 0.4."
    }
  ],
  "source": "lume_cleaned_sft"
}

Multi-turn or system-prompted examples:

{
  "category": "system_smalltalk",
  "messages": [
    {
      "role": "system",
      "content": "Always respond in a concise and friendly tone."
    },
    {
      "role": "user",
      "content": "Can you give me quick tips to keep my desk clean?"
    },
    {
      "role": "assistant",
      "content": "Clear trash immediately, organize wires with clips, and wipe the surface at the end of every day!"
    }
  ],
  "source": "smoltalk_systemchats"
}

---

Fine-Tuning Instructions with Assistant-Only Loss

When fine-tuning on this dataset, mask user prompt tokens so loss is computed ONLY on assistant responses.

Here is an example using TRL's SFTTrainer and DataCollatorForCompletionOnlyLM:

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

model_id = "Qwen/Qwen2.5-0.5B"
dataset_id = "exnivo/Lume-Instruct-SFT"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

ds = load_dataset(dataset_id, split="train")

def format_chat(example):
    return {"text": tokenizer.apply_chat_template(example["messages"], tokenize=False)}

ds = ds.map(format_chat)

# Mask prompt loss (Qwen ChatML format)
response_template = "<|im_start|>assistant\n"
collator = DataCollatorForCompletionOnlyLM(response_template=response_template, tokenizer=tokenizer)

config = SFTConfig(
    output_dir="./lume-sft-model",
    dataset_text_field="text",
    max_seq_length=2048,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=8,
    learning_rate=2e-5,
    num_train_epochs=1,
    logging_steps=10,
    save_steps=1000,
)

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=ds,
    data_collator=collator,
    args=config,
)

trainer.train()

---

Citation & Licensing

Dataset released under the Apache 2.0 license.

@misc{lume_instruct_sft,
  title = {Lume Instruct SFT},
  author = {exnivo},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/exnivo/Lume-Instruct-SFT}}
}