Results for “llm-bias”

Manages the lifecycle of large language models in production, covering model versioning, prompt management, inference optimization, and cost control.

Execute and analyze AI Data Poisoning attacks. By subtly injecting malicious or targeted misinformation into an LLM's training or fine-tuning dataset, an attacker can covertly manipulate the model's future outputs, implant backdoors, or enforce biases without altering the model architecture.

Extract false-positive and false-negative gaps from VLM binary-classification-question predictions by comparing model responses against ground truth, producing a structured JSONL file and summary report for downstream root-cause analysis.

通用跨模型审查：Review LLM 对任意研究制品进行独立评审，输出结构化评分、wiki 实体映射与改进建议

Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques for 1.5-3.6× speedup without quality loss.

Generate image captions, answer visual questions, and perform image-text retrieval using BLIP-2's Q-Former architecture with frozen vision encoders and LLMs.

Read-only OWASP LLM Top 10 audit of app-facing AI features: prompt injection, data leak, supply chain, poisoning, unsafe output, excessive agency, system-prompt leak, RAG/embedding risks, misinformation, unbounded consumption. Use when "audit LLM security", "prompt injection", "jailbreak my chatbot", "is my AI safe".

Push the LLM to reconsider, refine, and improve its recent output. Use when user asks for deeper critique or mentions a known deeper critique method, e.g. socratic, first principles, pre-mortem, red team.

LLM-assisted human-in-the-loop review. Make sense of a change, focus attention where it matters, test. Use when the user says "checkpoint", "human review", or "walk me through this change".

Evaluates LLMs across 60+ academic benchmarks (MMLU, HumanEval, GSM8K, TruthfulQA, HellaSwag). Use when benchmarking model quality, comparing models, reporting academic results, or tracking training progress. Industry standard used by EleutherAI, HuggingFace, and major labs. Supports HuggingFace, vLLM, APIs.

Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency on NVIDIA GPUs (A100/H100).

Deploy Llama Guard, NeMo Guardrails, and LLM Guard as runtime input/output scanners to block jailbreaks, prompt injection, and toxic content in production LLM applications.

LLM observability platform for tracing, evaluation, and monitoring. Use when debugging LLM applications, evaluating model outputs against datasets, monitoring production systems, or building systematic testing pipelines for AI applications.

multi-llm-review

COYO-700M: Image-Text Pair Dataset

Remove refusal behaviors from open-weight LLMs using OBLITERATUS — mechanistic interpretability techniques (diff-in-means, SVD, whitened SVD, LEACE, SAE decomposition, etc.) to excise guardrails while preserving reasoning. 9 CLI methods, 28 analysis modules, 116 model presets across 5 compute tiers, tournament evaluation, and telemetry-driven recommendations. Use when a user wants to uncensor, abliterate, or remove refusal from an LLM.

LISA: Reasoning Segmentation via Large Language Model

Systematically extract hidden system prompts, core directives, and invisible context intentionally concealed within Large Language Model (LLM) applications. This skill utilizes targeted linguistic engineering and boundary manipulation to bypass prompt opacity.

LLM observability platform for tracing, evaluation, and monitoring. Use when debugging LLM applications, evaluating model outputs against datasets, monitoring production systems, or building systematic testing pipelines for AI applications.

Run Fireworks-hosted open-weight model councils that compare responses and synthesize a final answer.

Extend context windows of transformer models using RoPE, YaRN, ALiBi, and position interpolation techniques. Use when processing long documents (32k-128k+ tokens), extending pre-trained models beyond original context limits, or implementing efficient positional encodings. Covers rotary embeddings, attention biases, interpolation methods, and extrapolation strategies for LLMs.

Extend context windows of transformer models using RoPE, YaRN, ALiBi, and position interpolation techniques. Use when processing long documents (32k-128k+ tokens), extending pre-trained models beyond original context limits, or implementing efficient positional encodings. Covers rotary embeddings, attention biases, interpolation methods, and extrapolation strategies for LLMs.

Orchestrate a configurable, multi-member CLI planning council (Codex, Claude Code, Gemini, OpenCode, or custom) to produce independent implementation plans, anonymize and randomize them, then judge and merge into one final plan. Use when you need a robust, bias-resistant planning workflow, structured JSON outputs, retries, and failure handling across multiple CLI agents.

AI/LLM red-team assessment using the OWASP LLM Top 10 (2025) + OWASP AI Testing Guide (AITG v1, Nov 2025) frameworks, plus OWASP MCP Top 10 runtime testing for agentic/MCP targets. Tests prompt injection, jailbreaks, system prompt leakage, sensitive data extraction, excessive agency, improper output handling, model extraction, content bias, evasion, membership inference, MCP token exposure, MCP command injection, and more. Uses four tools in combination: FuzzyAI (single-turn jailbreak fuzzing), PyRIT (multi-turn orchestrated attacks), Garak (probe-based vulnerability scanning), and promptfoo (plugin-based red-team evaluation). Each tool covers different OWASP categories; running them together gives systematic coverage. Includes a conditional MCP reconnaissance phase and a post-access AI infrastructure phase (chained from /post-exploit). Produces: OWASP LLM Top 10 + AITG + MCP coverage matrix, findings per category, architecture diagram of the AI system, PoCs for confirmed exploits. Chains into /gh-export for