Sam Joshua

SamJoshua

·

https://x.com/SAM_JOSHUA_S

AI & ML interests

AI agents, LLM tool use, agent security, evals, multi-agent systems

Recent Activity

new activity 5 days ago

swiss-ai/Apertus-8B-Instruct-2509:Sync chat_template.jinja with vLLM's Apertus tool template so tools render in transformers

updated a Space 5 days ago

SamJoshua/agent-harness-checks

new activity 5 days ago

ibm-granite/granite-4.1-3b:chat_template: tool_call.name is concatenated into JSON without escaping — name rewrite / key injection on re-render

View all activity

Organizations

New activity in swiss-ai/Apertus-8B-Instruct-2509 5 days ago

Sync chat_template.jinja with vLLM's Apertus tool template so tools render in transformers

#36 opened 5 days ago by

updated a Space 5 days ago

Agent Harness Checks

Silent failure checks for tool & computer-use agents

New activity in ibm-granite/granite-4.1-3b 5 days ago

chat_template: tool_call.name is concatenated into JSON without escaping — name rewrite / key injection on re-render

#7 opened 5 days ago by

New activity in nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1 5 days ago

expected_action calls tools not in `tools` (45 rows); 8 more have capitalized arg keys; README row count is wrong

#1 opened 5 days ago by

New activity in Qwen/Qwen3.8-27B 5 days ago

chat_template: OpenAI-style string tool_call.arguments still crashes with TypeError on |items

#240 opened 5 days ago by

New activity in openbmb/MiniCPM5-2B 5 days ago

chat_template: string tool-call arguments crash rendering, and non-string values are written as Python repr instead of JSON

#30 opened 5 days ago by

published a Space 5 days ago

Agent Harness Checks

Silent failure checks for tool & computer-use agents

New activity in microsoft/OmniParser-v2.0 5 days ago

`handler.py` checks for `"RBGA"` (typo) so RGBA screenshots never convert — plus Hub `intersection_area` still wrong

#38 opened 5 days ago by

New activity in patronus-studio/wolf-defender-prompt-injection 5 days ago

Mixed / int8_int4 ONNX exports are non-deterministic on CPU (same input → INJECTION or BENIGN)

#3 opened 5 days ago by

New activity in nvidia/Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1 5 days ago

Tool schemas use two different formats, and 3 records have target_args that can never trigger the verifier

#1 opened 5 days ago by

New activity in protectai/deberta-v3-base-prompt-injection-v2 5 days ago

Quick start truncates at 512 tokens: injections after that are never scored (tested, with a chunked fix)

#11 opened 5 days ago by

New activity in Qwen/Qwen2.5-Coder-1.5B-Instruct 5 days ago

Tested smolagents CodeAgent example for this model

#10 opened 5 days ago by

New activity in smolagents/aguvis-stage-2 5 days ago

Add dataset card: description, format, configs, and quick start

#3 opened 5 days ago by

New activity in smolagents/SmolVLM2-2.2B-Instruct-Agentic-GUI 5 days ago

Fix model card: base model, title, and image-text-to-text quick start

#3 opened 5 days ago by

updated a Space 6 days ago

Start Here

Sam Joshua - building AI agents

published a Space 6 days ago

Start Here

Sam Joshua - building AI agents

updated a Space 15 days ago

Sec RLVR Hub

Generate secure code and security advice with a defensive AI

published a Space 15 days ago

Sec RLVR Hub

Generate secure code and security advice with a defensive AI

updated a Space 17 days ago

Log Threat Classifier

published a Space 17 days ago

Log Threat Classifier