Dataset Viewer
Auto-converted to Parquet Duplicate
compression_arm
stringclasses
1 value
compression_scope
stringclasses
1 value
data_type
stringclasses
1 value
generation
dict
gold_answer
stringclasses
3 values
harvest_version
stringclasses
1 value
messages
listlengths
7
9
model
stringclasses
1 value
model_revision
stringclasses
1 value
schema_version
int64
2
2
segment_count
int64
27
29
source_dataset
stringclasses
1 value
sub_dataset
stringclasses
3 values
support_segment_ids
listlengths
3
5
task
stringclasses
3 values
task_id
stringclasses
3 values
tool_call_count
int64
3
5
tools
listlengths
1
1
verification
dict
compressed
input_segments
synthetic_expansion_agent
{ "max_tokens_per_turn": 512, "temperature": 0.1, "top_p": 0.95 }
Bay 37 | Bay 84 | Bay 82 | LATEST Bay 82
synthetic-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
27
synthetic-expansion-v2
latest_state
[ "seg_14", "seg_18", "seg_8" ]
According to all three movement records, list the bays assigned to asset AST-288-Z at 08:00, 11:00, and 15:00 in chronological order, then repeat the latest bay. Return exactly `FINAL: Bay A | Bay B | Bay C | LATEST Bay C`.
sea-565bba68-f8ec-54d7-8b9d-73a35c170c43
3
[ { "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, "expanded_support": [ "seg_14", "seg_18", "seg_8" ], "missing_support": [], "parsed_final": "FINAL: Bay 37 | Bay 84 | Bay 82 | LATEST Bay 82", "reason": "accepted" }
compressed
input_segments
synthetic_expansion_agent
{ "max_tokens_per_turn": 512, "temperature": 0.1, "top_p": 0.95 }
Western
synthetic-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
27
synthetic-expansion-v2
two_hop_join
[ "seg_25", "seg_14", "seg_21" ]
"In which region is the compliance facility assigned to the supplier of shipment SHP-655-G registere(...TRUNCATED)
sea-7f6364ef-281d-5e44-8940-8a6591576fd0
3
[{"function":{"description":"Return the original uncompressed text of one compressed context segment(...TRUNCATED)
{"accepted":true,"expanded_support":["seg_14","seg_21","seg_25"],"missing_support":[],"parsed_final"(...TRUNCATED)
compressed
input_segments
synthetic_expansion_agent
{ "max_tokens_per_turn": 512, "temperature": 0.1, "top_p": 0.95 }
delayed | released | cancelled | delayed | packed
synthetic-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
29
synthetic-expansion-v2
multi_key_lookup
[ "seg_16", "seg_19", "seg_26", "seg_23", "seg_7" ]
"Report the statuses of orders ORD-674-K, ORD-733-X, ORD-645-H, ORD-800-T, ORD-924-C in that order. (...TRUNCATED)
sea-a0ec50ca-f778-5787-bf55-d64625dbede1
5
[{"function":{"description":"Return the original uncompressed text of one compressed context segment(...TRUNCATED)
{"accepted":true,"expanded_support":["seg_16","seg_19","seg_23","seg_26","seg_7"],"missing_support":(...TRUNCATED)

Stage 3 Synthetic Expansion Agents — Verified Pilot

This is a small inspection pilot for native selective-expansion training. It is not the final-scale mixture.

Each initial user context contains 27–29 positional segments named seg_1, seg_2, and so on. Every segment contains its full source document wrapped in <|memory_start|>...<|memory_end|>. The assistant receives a native Qwen tool schema and may call expand({"segment_id": "seg_i"}); the matching tool result then places that same full document into the conversation. The Qwen teacher saw only short routing descriptions while generating calls; those descriptions are not stored as the train-time memory bodies.

Pilot results

  • Requested tasks: 5, one per task family
  • Verified accepted traces: 3
  • Native expansion calls: 11 (3–5 per trace)
  • Rejected traces: 2 (wrong exact final answers)
  • Model: Qwen/Qwen3-235B-A22B-Instruct-2507
  • Compression scope: initial input segments only
  • Per-segment floor: 512 whitespace-delimited words (882–972 Qwen tokens here)
  • Raw memory context: 25,627–27,328 Qwen tokens per accepted trace

The accepted split contains one latest-state task, one multi-key lookup, and one multi-hop join. The audit directory retains all five programmatic tasks and both rejected traces.

Training semantics

The tools field is passed to Qwen's native chat template, which renders the tool schema in its system tool block. Training loss applies to every assistant tool-call turn and the final assistant answer. System, user, and tool-result tokens are loss-masked. Memory bodies are extracted for the encoder/adapter path.

Verification

An accepted trace must use only native expand calls with a valid seg_i, receive the exact original text mapped to each call, expand every supporting segment, preserve call/result IDs, contain memory markers only in message text, and emit the exact requested FINAL: response.

Downloads last month
44