You can deploy this model and start earning money today!

Reality123b/
smollm3-3b-reasoning-engine

Name: smollm3-3b-reasoning-engine
Availability: InStock

Views

No views yet

Pure Reasoning Engine 🔍

A small, fast LLM fine-tuned on distilled frontier reasoning traces, conditioned to know nothing and treat the internet as its only source of truth.

Architecture

User Query → Decompose → Search → Filter → Reason → Answer with Citations

Component	Model	Size	Purpose
Reasoning Engine	SmolLM3-3B (fine-tuned)	3B params	Step-by-step reasoning over evidence
Epistemic Filter	DistilBERT (fine-tuned)	67M params	Pre-screens all retrieved text for quality
Internet Search	DuckDuckGo	Free API	Retrieves ground-truth evidence

Core Principle

"I know nothing. The internet is ground truth. My only job is to reason over what I retrieve."

The model NEVER guesses — it searches, retrieves, then reasons over verified sources.

Training Data

The reasoning model was fine-tuned on ~800K rows of distilled DeepSeek-R1 reasoning traces:

open-r1/OpenR1-Math-220k — Math reasoning (220K rows)
a-m-team/AM-DeepSeek-R1-Distilled-1.4M — Diversified reasoning (500K rows)
open-r1/codeforces-cots — Code reasoning (competitive programming)

The epistemic filter was trained on TurkuNLP/finerweb-10bt line-level quality labels.

Quick Start

bash

pip install -r requirements.txt
python inference_loop.py --query "What is the latest research on CRISPR?"

Interactive Mode

python inference_loop.py --interactive

Training

Reasoning Model (requires GPU with ≥24GB VRAM)

python train_reasoning_sft.py

Epistemic Filter (works on single GPU)

python train_epistemic_filter.py

Key Design Decisions

Small is Beautiful: 3B parameters is plenty when training is focused on reasoning
Humility Over Hubris: The model's strength is knowing it knows nothing
Reasoning > Memorization: Distill the skill of thinking, not the burden of facts
Defense First: The filter and source-critical training make the internet safe as ground truth
Cheap to Build: Base model is open-weight (Apache 2.0), training data is distilled, search is free

Model Links

Reasoning Engine: Reality123b/smollm3-3b-reasoning-engine
Epistemic Filter: Reality123b/epistemic-filter-v1

License

Apache 2.0 (same as SmolLM3 base model)

Reality123b/smollm3-3b-reasoning-engine