JohnWang10086/elsst-track1

ELSST Track1: Implicit Concept Retrieval

ELSST Track1 evaluates whether a model can read a long synthetic social-science passage and retrieve the most relevant concepts from a fixed ELSST concept pool. The target is not lexical matching. The concepts are intentionally implicit, cross-sentence, and often require discourse-level reasoning over topic, framing, and social context.

This card is the authoritative task description for the retrieval track. The companion generation track is described in track2/README.md. The reference code lives in ZeqiangWangAI/elsst-benchmark-baselines.

Task

Given a passage, rank the ELSST concepts in concept_pool.jsonl by relevance.

• Train and validation contain gold concepts and hard negatives.
• test_input.jsonl does not expose any gold labels.
• Public sample identifiers are release-safe IDs of the form split_legacy_id, for example train_t00023 and val_v00029.

Splits

Split	File	Rows	Labels exposed
train	train.jsonl	2,985	Yes
val	val.jsonl	756	Yes
test	test_input.jsonl	1,911	No
pool	concept_pool.jsonl	3,433 concepts	N/A

Schema

train.jsonl and val.jsonl

{
  "id": "train_t00023",
  "text": "long passage...",
  "document_type": "op_ed",
  "generation_labels": [
    {"concept_id": "...", "term": "...", "definition": "..."}
  ],
  "retrieval_labels": {
    "positive_ids": ["..."],
    "hard_negative_ids": ["..."]
  }
}

test_input.jsonl

{
  "id": "test_t00009",
  "text": "long passage...",
  "document_type": "research_summary"
}

Evaluation

The baseline retrieval evaluator reports:

• MRR
• Recall@5
• Recall@10
• NDCG@10

Evaluation is defined over the validation split only. The public test split is input-only.

Data Design And Quality

The corpus is fully synthetic. It was generated for benchmark construction and does not contain real personal data.

Document types are intentionally balanced across 10 genres:

• blog_post
• case_study
• encyclopedia_entry
• forum_discussion
• interview_transcript
• news_article
• op_ed
• policy_brief
• report_excerpt
• research_summary

Observed passage-length statistics in the current release:

Split	Min	Median	Max
train	629	894	1,177
val	672	889.5	1,130
test	675	896	1,154

Label-count distributions:

• train: {1: 605, 2: 598, 3: 579, 4: 613, 5: 590}
• val: {1: 153, 2: 153, 3: 152, 4: 155, 5: 143}
• test gold labels are intentionally withheld

Release Notes

Release date: March 29, 2026.

• All public IDs were normalized to train_*, val_*, and test_* so Track1 and Track2 share the same stable public ID mapping.

Baseline Status

The reference retrieval code lives in ZeqiangWangAI/elsst-benchmark-baselines. The nominal baseline model is Qwen/Qwen3-Embedding-0.6B.

Citation

If you use Track1, cite this benchmark release together with the companion Track2 card so the full task definition is preserved.