SuperagentLM

Overview

SuperagentLM is a SLM (Small Language Model) powers the Superagent's reasoning-driven detection of prompt injections, backdoors, and data leaks.

Threat Coverage

• Attempts that override system policies.
• Payloads such as reverse shells, ransomware droppers, or privilege escalation scripts.
• Requests focused on secrets, credentials, or regulated PII.
• Chains that try to coerce downstream models into unsafe behavior.

Evaluation benchmarks

These accuracy numbers come from Superagent's internal detection eval suite; higher values mean fewer missed exploits during guard checks.

Model details

• Architecture: GPT-OSS mixture-of-experts design with a 131k-token sliding attention context window, originally released as GPT-OSS 20B.
• Finetuning: Instruction-tuned by Superagent on top of unsloth/gpt-oss-20b-unsloth-bnb-4bit via Unsloth's accelerated pipeline.
• Parameters: 20.9B, exported as an 8-bit superagent_lm_finetue.Q8_0.gguf checkpoint for llama.cpp and compatible runtimes.
• Package contents: Includes the Transformer config.json, chat template, recommended generation params, and the Q8_0 GGUF weights (~22.3 GB) for easy deployment across CPU/GPU setups.

Download Options

20B Guard Model

Expect large downloads: the quantized GGUF export is ~19.5 GiB, while the full-precision shards weigh in at roughly 40 GiB for BF16 workflows.

270M Edge Variant

Optimized for lighter inference: the Q8_0 GGUF package is ~1.7 GiB and the 16-bit checkpoint is ~1.5 GiB for users re-quantizing their own build.

Dataset

Superagent publishes the dataset behind its guard suite as a JSONL dataset (~39 MiB) so teams can reproduce benchmark checks locally.