LFM3-1.7B-aura-grounded

Fine-tuned version of **Liquid AI's LFM3-1.7B** aligned with
**Direct Preference Optimization (DPO)** to eliminate
hallucinations in Retrieval-Augmented Generation (RAG) pipelines
and local agentic loops.

Small language models (< 3B) often fail to admit when retrieved
context does not contain the answer. When tested on unanswerable
retrieval queries, the base LFM3-1.7B model hallucinated over 94%
of the time, declining in only 3.4% of cases.

**LFM3-1.7B-aura-grounded** was trained with pure FP32 DPO on
dual Tesla T4 GPUs on grounded QA pairs and verified episode
preference data from the **Aura** runtime. It strictly adheres to
evidence: providing concise, accurate answers when context is
present, and cleanly declining with the designated fallback
sentence when information is absent.

---

## ⚡ Quickstart

Run it directly with Ollama:

```bash
ollama run milodule3/LFM3-1.7B-aura-grounded

Or call it via the Ollama Python library:

import ollama

response = ollama.chat(
    model="milodule3/LFM3-1.7B-aura-grounded",
    messages=[
        {
            "role": "user",
            "content": (
                "[1] Project Archimedes\n"
                "Archimedes is an edge alternator running local
                models at sub-second latencies.\n\n"
                "Question: What is Archimedes?"
            ),
        }
    ],
)
print(response["message"]["content"])

────── ## 🧩 Part of the Aura Ecosystem

This model is the lightweight grounding and memory-verification engine built for the Aura stack:

• aura-code https://github.com/DusanCar-sudo/aura-code (npm i -g aura-code): Autonomous terminal coding agent with persistent memory, interactive TUI, and multi-model routing. In Aura’s dual-tier architecture, this model serves as the local Archimedes alternator—verifying retrieved code context and memory before escalating heavy tasks to frontier models. • AgentMem https://agentmem-cloud.vercel.app: Shared cross-agent memory store linking agents across Linux, Windows, and cloud sessions. LFM3-1.7B-aura-grounded is specifically calibrated to read facts, decisions, and rules from agentmem without hallucinating or projecting stale context. • LeanProgressIQ: Research and development of sovereign, local- first agent tooling and grounded edge models. Learn more at leanproiq.com https://www.leanproiq.com. ────── ## 📊 Benchmark Results

Evaluated across training domains, completely unseen transfer domains, and live Aura episodes:

Metric / Evaluation Slice │ Base LF… │ LFM3-1.… │ Net Gain ────────────────────────────────┼──────────┼──────────┼──────────── Missing Evidence Refusal │ 3.4% │ 83.9% │ +80.5% (Training) │ │ │ Missing Evidence Refusal │ 5.8% │ 74.4% │ +68.6% (Unseen Domains) │ │ │ Answer Accuracy (Unseen │ 38.6% │ 47.4% │ +8.8% Transfer) │ │ │ Aura Retrieval Grounding │ 100.0% │ 100.0% │ 100% (Answerable) │ │ │ Correct Aura Retrieval Grounding │ 100.0% │ 100.0% │ 100% Clean (Unanswerable) │ │ │ Refusal ────── ## 🛠️ Prompt Format & System Contract

The model is pre-configured with the following grounding system prompt:

System: Answer the user's question using only the documents they
provide. Answer in the language of the question, in one or two
sentences. If the documents don't contain the answer, reply with
exactly this sentence and nothing else: I couldn't find the
answer in the provided documents.

[1] Document Title
<Document Text>

Question: <User Question>

When evidence is missing or insufficient, the model cleanly outputs: I couldn’t find the answer in the provided documents. ────── ## 🔗 Links & Resources

• CLI Agent: aura-code on npm https://www.npmjs.com/package/aura- code · GitHub Repository https://github.com/DusanCar-sudo/aura-code • Agent Memory: AgentMem Cloud https://agentmem-cloud.vercel.app • Hugging Face: LeanProgressIQ/LFM3-1.7B-aura-grounded https://huggingface.co/LeanProgressIQ/LFM3-1.7B-aura-grounded • Organization: leanproiq.com https://www.leanproiq.com