atenareply/lfm2.5-1.2b-noval-agentic

Model Overview

The atenareply/lfm2.5-1.2b-noval-agentic is a 1.2 billion parameter model developed by atenareply, specialized as an agentic maintenance assistant. It is the result of a Supervised Fine-Tuning (SFT) process using LoRA on the ParamΔ model, focusing on grounded tool-use trajectories. The model's primary domain is fictional Orbital Mining Corporation (OMC) technical documentation and Mars Express telemetry.

Key Capabilities

• Telemetry Anomaly Triage: Classifies alerts as routine or genuine with high precision (1.00) for genuine anomalies.
• Native Tool-Calling: Supports 8 specific tools, including get_channel_stats, query_telemetry, classify_alert, and schedule_maintenance, using a native LFM2 tool format.
• Autonomous Maintenance Loop: Capable of executing a full maintenance loop for genuine anomalies, involving consulting SOPs, issuing commands, scheduling maintenance, and opening incidents.
• Raw-Log Reading: Can process raw telemetry rows for complex tasks, demonstrating a difficulty-graded agentic curriculum.
• High Performance in Domain: Achieves a pass@1 of 0.782 for full trajectories and an action-loop accuracy of 0.782, significantly outperforming its ParamΔ baseline.

Good For

• Automated Operations: Ideal for use within a tool-execution loop where the model's outputs (tool calls) are executed, and results are fed back.
• Decision-Making Agents: Functions as the 'decision brain' for complex, multi-step maintenance and triage processes.
• Verifiable Agentic Systems: Benefits from innovations like a deterministic verifier for evaluation and reward, ensuring high reliability in its specific domain.

Limitations

• Requires Tool Execution Loop: The model must be used within a multi-turn loop that executes tools and feeds results back; it will hallucinate or stall in plain chat without tool execution.
• Fictional Domain: Primarily trained on fictional data, meaning real-world applicability outside the OMC/Mars-Express context may vary.
• Recall Rate: Has a recall of 0.63 for genuine anomalies, indicating approximately 37% of genuine issues might be missed, though with perfect precision (no false alarms).