compression_arm string | compression_scope string | data_type string | generation dict | gold_answer string | harvest_version string | messages list | model string | model_revision string | schema_version int64 | segment_count int64 | source_dataset string | source_row_id string | sub_dataset string | support_segment_ids list | task string | task_id string | tool_call_count int64 | tools list | verification unknown |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | The National Football League Kickoff game marks the official start of the National Football League (NFL) regular season. This first game of the season is usually scheduled for the Thursday following Labor Day. | real-expansion-harvest-v1 | [
{
"content": "Answer the question using the supplied context segments.\nEach segment is named seg_i and initially appears through a compressed memory\nrepresentation rather than as ordinary plaintext. When exact information is\nmissing, call the expand tool with the required segment_id to place that\nsegment's ... | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | PrimeQA/clapnq | -5815149352480677527 | grounded_qa | [
"seg_7"
] | when is the first nfl football game of the season
Return exactly `FINAL: ANSWER`, replacing ANSWER with your answer. | rea-8f8d483cd4436d2f1888e047 | 1 | [
{
"function": {
"description": "Return the original uncompressed text of one compressed context segment. Use this when exact details are missing from a segment summary.",
"name": "expand",
"parameters": {
"additionalProperties": false,
"properties": {
"segment_id": {
... | {
"accepted": true,
"answer_metric": "clapnq_containment_or_token_f1",
"answer_score": 0.6000000000000001,
"expanded_support": [
"seg_7"
],
"missing_support": [],
"parsed_final": "The first NFL football game of the season is usually played on the Thursday following Labor Day.",
"reason": "accepted:c... |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | yes | real-expansion-harvest-v1 | [
{
"content": "Answer the question using the supplied context segments.\nEach segment is named seg_i and initially appears through a compressed memory\nrepresentation rather than as ordinary plaintext. When exact information is\nmissing, call the expand tool with the required segment_id to place that\nsegment's ... | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | qiaojin/PubMedQA:pqa_labeled | 23361217 | health | [
"seg_1"
] | Does the type of tibial component affect mechanical alignment in unicompartmental knee replacement?
Return exactly `FINAL: ANSWER`, replacing ANSWER with your answer. | rea-29f95431e9c4cc92a7191e71 | 1 | [
{
"function": {
"description": "Return the original uncompressed text of one compressed context segment. Use this when exact details are missing from a segment summary.",
"name": "expand",
"parameters": {
"additionalProperties": false,
"properties": {
"segment_id": {
... | {
"accepted": true,
"answer_metric": "pubmedqa_decision",
"expanded_support": [
"seg_1"
],
"missing_support": [],
"parsed_final": "Yes, the type of tibial component affects mechanical alignment in unicompartmental knee replacement. The onlay (metal-backed) group achieved significantly better mechanical ... |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | Patty and Selma Bouvier are Marge Simpson's older twin sisters. | real-expansion-harvest-v1 | [
{
"content": "Answer the question using the supplied context segments.\nEach segment is named seg_i and initially appears through a compressed memory\nrepresentation rather than as ordinary plaintext. When exact information is\nmissing, call the expand tool with the required segment_id to place that\nsegment's ... | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | PrimeQA/clapnq | -4770500272958436225 | grounded_qa | [
"seg_7"
] | what are the names of marge's sisters
Return exactly `FINAL: ANSWER`, replacing ANSWER with your answer. | rea-659bf5d3ac3a6d9a9a8a086e | 1 | [
{
"function": {
"description": "Return the original uncompressed text of one compressed context segment. Use this when exact details are missing from a segment summary.",
"name": "expand",
"parameters": {
"additionalProperties": false,
"properties": {
"segment_id": {
... | {
"accepted": true,
"answer_metric": "clapnq_containment_or_token_f1",
"answer_score": 0.4285714286,
"expanded_support": [
"seg_7"
],
"missing_support": [],
"parsed_final": "Patty and Selma",
"reason": "accepted:clapnq_containment_or_token_f1"
} |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | NotMentioned | real-expansion-harvest-v1 | [
{
"content": "Answer the question using the supplied context segments.\nEach segment is named seg_i and initially appears through a compressed memory\nrepresentation rather than as ordinary plaintext. When exact information is\nmissing, call the expand tool with the required segment_id to place that\nsegment's ... | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | stanfordnlp/contract-nli | 177::nda-12 | legal | [
"seg_6"
] | Does the contract entail, contradict, or not mention this statement: Receiving Party may independently develop information similar to Confidential Information.
Return exactly `FINAL: ANSWER`, replacing ANSWER with your answer. | rea-cffd968901dedad485a2080c | 1 | [
{
"function": {
"description": "Return the original uncompressed text of one compressed context segment. Use this when exact details are missing from a segment summary.",
"name": "expand",
"parameters": {
"additionalProperties": false,
"properties": {
"segment_id": {
... | {
"accepted": true,
"answer_metric": "contract_nli_label",
"expanded_support": [
"seg_6"
],
"missing_support": [],
"parsed_final": "not mention",
"reason": "accepted:contract_nli_label"
} |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | Entailment | real-expansion-harvest-v1 | [{"content":"Answer the question using the supplied context segments.\nEach segment is named seg_i a(...TRUNCATED) | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | stanfordnlp/contract-nli | 245::nda-5 | legal | [
"seg_3"
] | "Does the contract entail, contradict, or not mention this statement: Receiving Party may share some(...TRUNCATED) | rea-bf6a718845661e76a0a8d94d | 1 | [{"function":{"description":"Return the original uncompressed text of one compressed context segment(...TRUNCATED) | {"accepted":true,"answer_metric":"contract_nli_label","expanded_support":["seg_3"],"missing_support"(...TRUNCATED) |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | NotMentioned | real-expansion-harvest-v1 | [{"content":"Answer the question using the supplied context segments.\nEach segment is named seg_i a(...TRUNCATED) | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | stanfordnlp/contract-nli | 154::nda-18 | legal | [
"seg_1"
] | "Does the contract entail, contradict, or not mention this statement: Receiving Party shall not soli(...TRUNCATED) | rea-5487c93019b6f3a8af1d28bb | 1 | [{"function":{"description":"Return the original uncompressed text of one compressed context segment(...TRUNCATED) | {"accepted":true,"answer_metric":"contract_nli_label","expanded_support":["seg_1"],"missing_support"(...TRUNCATED) |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | 6.0% | real-expansion-harvest-v1 | [{"content":"Answer the question using the supplied context segments.\nEach segment is named seg_i a(...TRUNCATED) | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | bevaya/FinQA | ETR/2016/page_17.pdf-1 | finance | [
"seg_8"
] | "what is the growth rate in net revenue in 2016?\nReturn exactly `FINAL: ANSWER`, replacing ANSWER w(...TRUNCATED) | rea-eebee0df10d4b6551f37006f | 1 | [{"function":{"description":"Return the original uncompressed text of one compressed context segment(...TRUNCATED) | {"accepted":true,"answer_metric":"finqa_numeric","expanded_support":["seg_8"],"missing_support":[],"(...TRUNCATED) |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | 3.52 | real-expansion-harvest-v1 | [{"content":"Answer the question using the supplied context segments.\nEach segment is named seg_i a(...TRUNCATED) | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | bevaya/FinQA | KIM/2009/page_145.pdf-1 | finance | [
"seg_1"
] | "in 2009 what was the ratio of the interest rate caps to swaps\nReturn exactly `FINAL: ANSWER`, repl(...TRUNCATED) | rea-8ec8103ba29613969f238e0b | 1 | [{"function":{"description":"Return the original uncompressed text of one compressed context segment(...TRUNCATED) | {"accepted":true,"answer_metric":"finqa_numeric","expanded_support":["seg_1"],"missing_support":[],"(...TRUNCATED) |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | yes | real-expansion-harvest-v1 | [{"content":"Answer the question using the supplied context segments.\nEach segment is named seg_i a(...TRUNCATED) | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | qiaojin/PubMedQA:pqa_labeled | 12607666 | health | [
"seg_3"
] | "Is extended aortic replacement in acute type A dissection justifiable?\nReturn exactly `FINAL: ANSW(...TRUNCATED) | rea-dedbf3a2224f1976d6728b47 | 1 | [{"function":{"description":"Return the original uncompressed text of one compressed context segment(...TRUNCATED) | {"accepted":true,"answer_metric":"pubmedqa_decision","expanded_support":["seg_3"],"missing_support":(...TRUNCATED) |
compressed | input_segments | real_expansion_agent | {
"max_tokens_per_turn": 768,
"temperature": 0,
"top_p": 0.95
} | Entailment | real-expansion-harvest-v1 | [{"content":"Answer the question using the supplied context segments.\nEach segment is named seg_i a(...TRUNCATED) | Qwen/Qwen3-235B-A22B-Instruct-2507 | ac9c66cc9b46af7306746a9250f23d47083d689e | 2 | 8 | stanfordnlp/contract-nli | 187::nda-12 | legal | [
"seg_6"
] | "Does the contract entail, contradict, or not mention this statement: Receiving Party may independen(...TRUNCATED) | rea-50d92c0ec06be2500ff1ec2d | 1 | [{"function":{"description":"Return the original uncompressed text of one compressed context segment(...TRUNCATED) | {"accepted":true,"answer_metric":"contract_nli_label","expanded_support":["seg_6"],"missing_support"(...TRUNCATED) |
Stage 3 Real-Source Expansion Agents — Pilot
This inspection pilot converts pinned training examples from real legal, financial, biomedical, and grounded-QA corpora into native selective-expansion traces. It is not the final-scale mixture.
Each row contains eight positional seg_i blocks. Every initial segment holds
512–896 words of real source material wrapped in
<|memory_start|>...<|memory_end|>. Qwen3-235B-A22B-Instruct-2507 receives a
native expand({"segment_id": "seg_i"}) tool, expands a source segment, and
answers using the returned plaintext.
Results
- Source tasks: 20 (four from each source)
- Accepted traces: 10
- Rejected traces retained in
audit/rejected.jsonl: 10 - Segments in the source task set: 160
- Total real-source words: 116,768
- Accepted sources: ContractNLI 4, FinQA 2, PubMedQA 2, CLAPNQ 2
- Native expansion calls per accepted trace: 1
- Model revision:
ac9c66cc9b46af7306746a9250f23d47083d689e
Verification
All accepted traces must use a valid native tool call, receive the exact source
text for that seg_i, and expand every annotated support segment. Answers use
the source's native metric: ContractNLI canonical labels, FinQA numeric
equivalence, PubMedQA decision labels, and CLAPNQ grounded containment/token-F1.
Strictly wrong, missing-final, and tool-limit traces are excluded from train.
Provenance
metadata/source-manifest.json pins all upstream dataset revisions. Licensing
continues to follow each upstream source; users should review those terms before
redistribution or commercial use.
- Downloads last month
- 36