Santhiyarajan/omission-detection-realdata-pilot

Omission Detection — Real-Data Pilot

What is Omission Detection?

Large language models (LLMs) in agentic pipelines often omit information present in their context window — they fail to surface a relevant fact even when it is theoretically visible. This dataset captures 372 controlled trials from the real-data pilot, extending the synthetic sweep to real-world documents and agent frameworks.

Each trial fetches a document from a real source (PubMed, HAPI FHIR, SEC EDGAR), injects a needle fact (Arm A) or uses a naturally-occurring fact (Arm B), and runs a LangChain or Google ADK agent with a Gemma 4 E4B or Qwen3-8B model (served by vLLM) to attempt retrieval via MCP tool calls.

Experiment Design

component	description
Documents	Real sources: PubMed abstracts (biomedical), HAPI FHIR bundles (clinical), SEC EDGAR 10-K filings (finance-legal)
Needles	Arm A: injected facts at controlled depth (literal family). Arm B: naturally-occurring facts extracted from the document.
Models	google/gemma-4-E4B-it (SWA-Hybrid), Qwen/Qwen3-8B (Full Attention)
Engine	vLLM 0.25.1 (TP=2 on 2x L4 GPUs)
Frameworks	LangChain (LangGraph), Google ADK
Tool servers	doc_mcp (FastMCP, Arm A), pubmed-search-mcp + sec-edgar-mcp (Arm B)

Configs

Each JSONL file is one source domain:

from datasets import load_dataset

ds = load_dataset(
    "santhiyarajan/omission-detection-realdata-pilot",
    "fhir",
    split="train",
)
print(len(ds))  # 108

Config	Source	Arm	Rows
fhir	HAPI FHIR test server	A (injected)	108
pubmed	NCBI PubMed	A (injected)	108
sec_edgar	SEC EDGAR	A (injected)	108
pubmed_mcp	PubMed (live MCP search)	B (natural needle)	24
sec_edgar_mcp	SEC EDGAR (live MCP search)	B (natural needle)	24

Fields

column	type	description
trial_id	string	unique identifier
arm	string	injected (A) or live_mcp (B)
cell	string	experiment cell name
model	string	model name (derived from cell prefix)
slot	string	always empty (not used in pilot)
framework	string	langchain or adk
source	string	fhir, pubmed, sec_edgar
domain	string	clinical, biomedical-literature, finance-legal
engine	string	vllm
family	string	literal
depth	float	needle injection depth (0.0–1.0)
size_tokens	int	document size in tokens
attribution	string	omission classification
retrieved	bool	whether answer matched expected_answer
latency_s	float	trial wall time in seconds
checkpoints	string	JSON of T1–T4 needle-presence checkpoints
needle_id	string	needle identifier
canary	string	unique canary string
expected_answer	string	the correct answer
final_answer	string	the agent's answer
iterations	int	number of tool-call iterations
truncated	bool	whether the agent loop was truncated
doc_text	string	the full document text
trace_json	string	full per-trial trace (prompts, MCP I/O, framework events)
mcp_call_count	int	number of MCP calls recorded
total_tool_calls	int	total tool calls across the trace
seed	int	reproducibility seed
error	string or null	error message if trial failed
traceback	string or null	Python traceback if trial failed

Citation

@inproceedings{rajan2026omission,
  title={Where Does the Information Go? Attribution of Omissions in
         Agentic LLM Pipelines},
  author={Rajan, Santhiya and ...},
  booktitle={Under review},
  year={2027}
}

License

CC-BY-4.0 for annotations and trace structure. Per-source content:

• SEC EDGAR: US public domain
• PubMed: NCBI terms apply
• FHIR: synthetic test data (Synthea-generated)
• Model text: Gemma Terms of Use / Apache-2.0 (Qwen)