id stringclasses 8
values | timestamp stringclasses 8
values | agent stringclasses 6
values | tool_name stringclasses 8
values | transform stringclasses 4
values | content_type stringclasses 7
values | query_context stringclasses 8
values | raw_tokens int64 1.16k 6.4k | compressed_tokens int64 159 540 | tokens_saved int64 794 6.06k | savings_pct float64 64.4 94.8 | compression_ratio float64 0.05 0.36 | raw_chars int64 3.16k 25.6k | compressed_chars int64 403 1.98k | moe_primary_expert stringclasses 4
values | moe_secondary_expert stringclasses 3
values | moe_expert_weights dict | moe_features dict | projection_certificate dict | iroh_ticket stringclasses 1
value | iroh_blake3_hash stringclasses 8
values | raw_content stringclasses 8
values | compressed_content stringclasses 8
values | sft_messages listlengths 3 3 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
04750bdbe56b39e65c3d5538 | 2026-10-30T17:37:47.752365+00:00 | crush-dev-local | bash:cargo_build | lean_context | build_log | Fix Rust compilation error in headroom-core router.rs | 1,385 | 493 | 892 | 64.4 | 0.356 | 5,542 | 1,975 | log_dedup_expert | lean_signal_expert | {
"log_dedup_expert": 0.4315,
"lean_signal_expert": 0.3238,
"json_schema_expert": 0.064,
"tabular_expert": 0.0625,
"diff_hunk_expert": 0.0624,
"code_skeleton_expert": 0.0558
} | {
"total_lines": 94,
"total_chars": 5542,
"json_density": 0.0053,
"tabular_density": 0,
"error_signal_density": 0.0957,
"diff_patch_density": 0,
"log_repetition_density": 0.5149,
"code_ast_density": 0,
"prose_entropy": 0.4851
} | {
"certificate_version": "1.1-sufficient-projection",
"projection_operator": "pi:log_dedup_expert",
"retained_subspace": [
"head_tail_4_lines",
"error_panic_windows",
"first_2_template_exemplars",
"run_length_cardinality"
],
"admissible_revisions": [
"R_failure_window_diagnosis",
"R_te... | blobapmrhmir2gewce3tpmrsseorcnfzg62brpbrww4dhnvuxqy3gn5vwyztr | f06ed879ba296788a8122b3cb5254d9b2419c854514f50e707ae0a110599afa8 | Compiling crate_dep_1 v0.4.1 (/workspace/crates/dep_1)
Compiling crate_dep_2 v0.4.2 (/workspace/crates/dep_2)
Compiling crate_dep_3 v0.4.3 (/workspace/crates/dep_3)
Compiling crate_dep_4 v0.4.4 (/workspace/crates/dep_4)
Compiling crate_dep_5 v0.4.5 (/workspace/crates/dep_5)
Compiling crate_dep_6 v0.4.... | Compiling crate_dep_1 v0.4.1 (/workspace/crates/dep_1)
Compiling crate_dep_2 v0.4.2 (/workspace/crates/dep_2)
Compiling crate_dep_3 v0.4.3 (/workspace/crates/dep_3)
... [33 lines dropped] ...
Compiling crate_dep_37 v0.4.7 (/workspace/crates/dep_37)
Compiling crate_dep_38 v0.4.8 (/workspace/crates/dep_38)... | [
{
"role": "system",
"content": "You are 8b.HeadroomZ, a lossless-signal context compression model for AI coding agents (opencode, crush, crush-dev-local, codewhale, claude, codex). Given a raw tool output, compress redundant boilerplate, passing logs, and middle array items while preserving all error signal... |
6354db3362fec7612b5c081d | 2026-10-30T17:37:47.783494+00:00 | opencode | bash:pytest | lean_context | traceback | Investigate InvalidStateError when a coalesced request is cancelled | 2,132 | 362 | 1,770 | 83.02 | 0.1698 | 8,530 | 1,448 | json_schema_expert | lean_signal_expert | {
"log_dedup_expert": 0.0657,
"lean_signal_expert": 0.3301,
"json_schema_expert": 0.4608,
"tabular_expert": 0.0478,
"diff_hunk_expert": 0.0414,
"code_skeleton_expert": 0.0542
} | {
"total_lines": 140,
"total_chars": 8530,
"json_density": 0.4821,
"tabular_density": 0,
"error_signal_density": 0.0964,
"diff_patch_density": 0,
"log_repetition_density": 0,
"code_ast_density": 0,
"prose_entropy": 0.5179
} | {
"certificate_version": "1.1-sufficient-projection",
"projection_operator": "pi:json_schema_expert",
"retained_subspace": [
"json_envelope_schema",
"head_3_tail_2_items",
"status_error_anomalies"
],
"admissible_revisions": [
"R_schema_inference",
"R_anomaly_triage",
"R_boundary_inspec... | blobapmrhmir2gewce3tpmrsseorcnfzg62brpbrww4dhnvuxqy3gn5vwyztr | 928c49ed932059349b6cd673bc8cfe49cd8823088e7bd50b603645bbcad3b5d1 | tests/test_proxy/test_suite_0.py::test_case_1 PASSED [ 0%]
tests/test_proxy/test_suite_0.py::test_case_2 PASSED [ 1%]
tests/test_proxy/test_suite_0.py::test_case_3 PASSED [ 2%]
tests/test_proxy/test_suite_0.py::test_case_4 PASSED [ 2%]
tests/test_proxy/test_suite_0.py::test_case_5 PASSED [ 3%]
tests/test_proxy/test_sui... | tests/test_proxy/test_suite_0.py::test_case_1 PASSED [ 0%]
tests/test_proxy/test_suite_0.py::test_case_2 PASSED [ 1%]
tests/test_proxy/test_suite_0.py::test_case_3 PASSED [ 2%]
... [76 lines dropped] ...
tests/test_proxy/test_suite_8.py::test_case_80 PASSED [ 57%]
tests/test_proxy/test_suite_8.py::test_case_81 PASSED [... | [
{
"role": "system",
"content": "You are 8b.HeadroomZ, a lossless-signal context compression model for AI coding agents (opencode, crush, crush-dev-local, codewhale, claude, codex). Given a raw tool output, compress redundant boilerplate, passing logs, and middle array items while preserving all error signal... |
cb5b92bb29f2ccbe8350287d | 2026-10-30T17:37:47.788703+00:00 | crush | git_diff | lean_context | git_diff | Inspect active_savings_percent denominator patch in server.py | 1,315 | 419 | 896 | 68.14 | 0.3186 | 5,239 | 1,654 | lean_signal_expert | diff_hunk_expert | {
"log_dedup_expert": 0.0828,
"lean_signal_expert": 0.5698,
"json_schema_expert": 0.083,
"tabular_expert": 0.0822,
"diff_hunk_expert": 0.1028,
"code_skeleton_expert": 0.0795
} | {
"total_lines": 119,
"total_chars": 5239,
"json_density": 0.0126,
"tabular_density": 0,
"error_signal_density": 0,
"diff_patch_density": 0.0504,
"log_repetition_density": 0,
"code_ast_density": 0,
"prose_entropy": 0.9496
} | {
"certificate_version": "1.1-sufficient-projection",
"projection_operator": "pi:lean_signal_expert",
"retained_subspace": [
"head_tail_anchors",
"error_traceback_windows",
"grep_match_collars"
],
"admissible_revisions": [
"R_error_localization",
"R_exit_status_triage",
"R_stack_frame_... | blobapmrhmir2gewce3tpmrsseorcnfzg62brpbrww4dhnvuxqy3gn5vwyztr | 193d985fb8d76ecc5ca671bff40cc7614130f0cc876e5117e8653d7841910d2f | context_line_1: unchanged workspace state
context_line_2: unchanged workspace state
context_line_3: unchanged workspace state
context_line_4: unchanged workspace state
context_line_5: unchanged workspace state
context_line_6: unchanged workspace state
context_line_7: unchanged workspace state
context_line_8: unchanged ... | context_line_1: unchanged workspace state
context_line_2: unchanged workspace state
context_line_3: unchanged workspace state
... [26 lines dropped] ...
context_line_30: unchanged workspace state
context_line_31: unchanged workspace state
context_line_32: unchanged workspace state
context_line_33: unchanged workspace s... | [
{
"role": "system",
"content": "You are 8b.HeadroomZ, a lossless-signal context compression model for AI coding agents (opencode, crush, crush-dev-local, codewhale, claude, codex). Given a raw tool output, compress redundant boilerplate, passing logs, and middle array items while preserving all error signal... |
71b837fdb1b21c99e8dd9d7b | 2026-10-30T17:37:47.803952+00:00 | codewhale | mcp:query_production_events | smart_crusher | json_array | Find why auth-token-verifier timed out in eu-central-1 | 6,398 | 335 | 6,063 | 94.76 | 0.0524 | 25,595 | 1,343 | json_schema_expert | lean_signal_expert | {
"log_dedup_expert": 0.1224,
"lean_signal_expert": 0.1597,
"json_schema_expert": 0.7014,
"tabular_expert": 0.0064,
"diff_hunk_expert": 0.0047,
"code_skeleton_expert": 0.0054
} | {
"total_lines": 1197,
"total_chars": 25595,
"json_density": 1,
"tabular_density": 0,
"error_signal_density": 0.558,
"diff_patch_density": 0,
"log_repetition_density": 0.5577,
"code_ast_density": 0,
"prose_entropy": 0
} | {
"certificate_version": "1.1-sufficient-projection",
"projection_operator": "pi:json_schema_expert",
"retained_subspace": [
"json_envelope_schema",
"head_3_tail_2_items",
"status_error_anomalies"
],
"admissible_revisions": [
"R_schema_inference",
"R_anomaly_triage",
"R_boundary_inspec... | blobapmrhmir2gewce3tpmrsseorcnfzg62brpbrww4dhnvuxqy3gn5vwyztr | aa8739ffe2a48787bc355b5991daa2384dc3414ec411e434aa2889a048ae22f7 | {
"events": [
{
"id": "evt_0001",
"service": "payment-gateway",
"status": "ok",
"latency_ms": 13,
"error": null,
"region": "eu-central-1"
},
{
"id": "evt_0002",
"service": "payment-gateway",
"status": "ok",
"latency_ms": 14,
"error": null,
... | {
"events": [
{
"id": "evt_0001",
"service": "payment-gateway",
"status": "ok",
"latency_ms": 13,
"error": null,
"region": "eu-central-1"
},
{
"id": "evt_0002",
"service": "payment-gateway",
"status": "ok",
"latency_ms": 14,
"error": null,
... | [
{
"role": "system",
"content": "You are 8b.HeadroomZ, a lossless-signal context compression model for AI coding agents (opencode, crush, crush-dev-local, codewhale, claude, codex). Given a raw tool output, compress redundant boilerplate, passing logs, and middle array items while preserving all error signal... |
2305262e233051886e452dbf | 2026-10-30T17:37:47.813718+00:00 | claude | bash:npm_run_build | lean_context | build_log | Fix TypeScript build error in ConstellationHud.tsx | 1,812 | 368 | 1,444 | 79.69 | 0.2031 | 7,248 | 1,472 | log_dedup_expert | lean_signal_expert | {
"log_dedup_expert": 0.5372,
"lean_signal_expert": 0.2429,
"json_schema_expert": 0.0563,
"tabular_expert": 0.0563,
"diff_hunk_expert": 0.0563,
"code_skeleton_expert": 0.051
} | {
"total_lines": 109,
"total_chars": 7248,
"json_density": 0,
"tabular_density": 0,
"error_signal_density": 0.0826,
"diff_patch_density": 0,
"log_repetition_density": 0.6083,
"code_ast_density": 0,
"prose_entropy": 0.3917
} | {
"certificate_version": "1.1-sufficient-projection",
"projection_operator": "pi:log_dedup_expert",
"retained_subspace": [
"head_tail_4_lines",
"error_panic_windows",
"first_2_template_exemplars",
"run_length_cardinality"
],
"admissible_revisions": [
"R_failure_window_diagnosis",
"R_te... | blobapmrhmir2gewce3tpmrsseorcnfzg62brpbrww4dhnvuxqy3gn5vwyztr | 1c80c9c8c03a389f9844f88c0252098447bfd1d1c2e27b08bacddeac14885cf5 | info - Compiling route /api/v1/endpoint-1 (validating schema...)
info - Compiling route /api/v1/endpoint-2 (validating schema...)
info - Compiling route /api/v1/endpoint-3 (validating schema...)
info - Compiling route /api/v1/endpoint-4 (validating schema...)
info - Compiling route /api/v1/endpoint-5 (validating s... | info - Compiling route /api/v1/endpoint-1 (validating schema...)
info - Compiling route /api/v1/endpoint-2 (validating schema...)
info - Compiling route /api/v1/endpoint-3 (validating schema...)
... [46 lines dropped] ...
info - Compiling route /api/v1/endpoint-50 (validating schema...)
info - Compiling route /api... | [
{
"role": "system",
"content": "You are 8b.HeadroomZ, a lossless-signal context compression model for AI coding agents (opencode, crush, crush-dev-local, codewhale, claude, codex). Given a raw tool output, compress redundant boilerplate, passing logs, and middle array items while preserving all error signal... |
c3142308f6036c714d2cc2ad | 2026-10-30T17:37:47.841643+00:00 | headroomz | tool_output | lean_context | tool_output | 5,000 | 540 | 4,460 | 89.2 | 0.108 | 3,164 | 403 | log_dedup_expert | lean_signal_expert | {
"log_dedup_expert": 0.5899,
"lean_signal_expert": 0.1831,
"json_schema_expert": 0.0571,
"tabular_expert": 0.0571,
"diff_hunk_expert": 0.0571,
"code_skeleton_expert": 0.0558
} | {
"total_lines": 250,
"total_chars": 3164,
"json_density": 0,
"tabular_density": 0,
"error_signal_density": 0.018,
"diff_patch_density": 0,
"log_repetition_density": 0.6448,
"code_ast_density": 0,
"prose_entropy": 0.3552
} | {
"certificate_version": "1.1-sufficient-projection",
"projection_operator": "pi:log_dedup_expert",
"retained_subspace": [
"head_tail_4_lines",
"error_panic_windows",
"first_2_template_exemplars",
"run_length_cardinality"
],
"admissible_revisions": [
"R_failure_window_diagnosis",
"R_te... | blobapmrhmir2gewce3tpmrsseorcnfzg62brpbrww4dhnvuxqy3gn5vwyztr | 4f98fe21a7865461d467507b9e4d9f5a05893b89ce0af120f09e47cbb799e695 | log line 0
log line 1
log line 2
log line 3
log line 4
log line 5
log line 6
log line 7
log line 8
log line 9
log line 10
log line 11
log line 12
log line 13
log line 14
log line 15
log line 16
log line 17
log line 18
log line 19
log line 20
log line 21
log line 22
log line 23
log line 24
log line 25
log line 26
log li... | log line 0
log line 1
log line 2
... [37 lines dropped] ...
log line 40
log line 41
log line 42
log line 43
log line 44
log line 45
log line 46
log line 47
log line 48
log line 49
error: compilation failed at line 50
log line 51
log line 52
log line 53
log line 54
log line 55
log line 56
log line 57
log line 58
log lin... | [
{
"role": "system",
"content": "You are 8b.HeadroomZ, a lossless-signal context compression model for AI coding agents (opencode, crush, crush-dev-local, codewhale, claude, codex). Given a raw tool output, compress redundant boilerplate, passing logs, and middle array items while preserving all error signal... | |
f274d7a936d2dffd59930456 | 2026-10-30T17:37:47.979132+00:00 | crush-dev-local | mcp:postgres_query_shards | moe_optimizer:tabular_expert | markdown_table | Identify which PostgreSQL shard has a p99 latency spike | 1,512 | 159 | 1,353 | 89.48 | 0.1052 | 6,049 | 637 | tabular_expert | lean_signal_expert | {
"log_dedup_expert": 0.0124,
"lean_signal_expert": 0.0226,
"json_schema_expert": 0.0124,
"tabular_expert": 0.9276,
"diff_hunk_expert": 0.0124,
"code_skeleton_expert": 0.0124
} | {
"total_lines": 96,
"total_chars": 6049,
"json_density": 0,
"tabular_density": 1,
"error_signal_density": 0,
"diff_patch_density": 0,
"log_repetition_density": 0,
"code_ast_density": 0,
"prose_entropy": 0
} | {
"certificate_version": "1.1-sufficient-projection",
"projection_operator": "pi:tabular_expert",
"retained_subspace": [
"column_header_schema",
"head_6_tail_4_rows",
"anomalous_error_rows"
],
"admissible_revisions": [
"R_column_schema_projection",
"R_head_tail_row_sampling",
"R_outlie... | blobapmrhmir2gewce3tpmrsseorcnfzg62brpbrww4dhnvuxqy3gn5vwyztr | 4bf292e4d83d30afc3a7561dd5be3800a463c7659437c93106e92008aa1f11ae | | shard_id | host | region | p99_ms | status | error_detail |
| --- | --- | --- | --- | --- | --- |
| shard_001 | fsn1-node-1 | eu-central | 12 | nominal | none |
| shard_002 | fsn1-node-2 | eu-central | 13 | nominal | none |
| shard_003 | fsn1-node-3 | eu-central | 14 | nominal | none |
| shard_004 | fsn1-node-4 | eu-... | | shard_id | host | region | p99_ms | status | error_detail |
| --- | --- | --- | --- | --- | --- |
| shard_001 | fsn1-node-1 | eu-central | 12 | nominal | none |
| shard_002 | fsn1-node-2 | eu-central | 13 | nominal | none |
| shard_003 | fsn1-node-3 | eu-central | 14 | nominal | none |
| shard_004 | fsn1-node-4 | eu-... | [
{
"role": "system",
"content": "You are 8b.HeadroomZ, a lossless-signal context compression model for AI coding agents (opencode, crush, crush-dev-local, codewhale, claude, codex). Given a raw tool output, compress redundant boilerplate, passing logs, and middle array items while preserving all error signal... |
6efb2763503c7514badb5d7c | 2026-10-30T17:37:47.988778+00:00 | opencode | read_file:router.rs | moe_optimizer:code_skeleton_expert | source_code | Inspect public API signatures of SovereignRouter | 1,156 | 362 | 794 | 68.69 | 0.3131 | 4,626 | 1,449 | lean_signal_expert | code_skeleton_expert | {
"log_dedup_expert": 0.0841,
"lean_signal_expert": 0.5638,
"json_schema_expert": 0.0974,
"tabular_expert": 0.0737,
"diff_hunk_expert": 0.0715,
"code_skeleton_expert": 0.1095
} | {
"total_lines": 97,
"total_chars": 4626,
"json_density": 0.0619,
"tabular_density": 0,
"error_signal_density": 0.0479,
"diff_patch_density": 0,
"log_repetition_density": 0.0207,
"code_ast_density": 0.1277,
"prose_entropy": 0.8723
} | {
"certificate_version": "1.1-sufficient-projection",
"projection_operator": "pi:lean_signal_expert",
"retained_subspace": [
"head_tail_anchors",
"error_traceback_windows",
"grep_match_collars"
],
"admissible_revisions": [
"R_error_localization",
"R_exit_status_triage",
"R_stack_frame_... | blobapmrhmir2gewce3tpmrsseorcnfzg62brpbrww4dhnvuxqy3gn5vwyztr | 9578165dc7863943e61ab3f49e81d6e426ef803640559bfcf19bb870bec70f7e | use serde::{Deserialize, Serialize};
use std::collections::BTreeMap;
pub struct SovereignRouter {
pub top_k: usize,
pub weights: [[f32; 7]; 6],
}
impl SovereignRouter {
pub fn new(top_k: usize) -> Self {
let internal_step_1 = (1usize * 3) ^ 0x5a;
let internal_step_2 = (2usize * 3) ^ 0x5a;
... | use serde::{Deserialize, Serialize};
use std::collections::BTreeMap;
... [65 lines dropped] ...
let internal_step_59 = (59usize * 3) ^ 0x5a;
let internal_step_60 = (60usize * 3) ^ 0x5a;
let internal_step_61 = (61usize * 3) ^ 0x5a;
let internal_step_62 = (62usize * 3) ^ 0x5a;
let... | [
{
"role": "system",
"content": "You are 8b.HeadroomZ, a lossless-signal context compression model for AI coding agents (opencode, crush, crush-dev-local, codewhale, claude, codex). Given a raw tool output, compress redundant boilerplate, passing logs, and middle array items while preserving all error signal... |
8b.HeadroomZ β Raw + Compressed Tool Output Training Traces (PeetPedro/headroomz-compression-traces)
Curated and live-streamed (raw_content, compressed_content, projection_certificate, sft_messages) pairs captured from 8b-is/headroomz across first-class coding harnesses (opencode, crush, crush-dev-local, codewhale, claude, and codex).
Dataset Summary
- Total Compression Pairs:
8 - Total Raw Tokens:
20,710 - Total Compressed Tokens:
3,038 - Total Tokens Saved:
17,672(85.33%average reduction) - TOIN Learned Tool Signatures:
75 - MOEOptimizer Specialist Routing (
tracel-ai/burnLinear[6,7]):lean_signal_expert,diff_hunk_expert,json_schema_expert,log_dedup_expert,tabular_expert,code_skeleton_expert - Formal Sufficiency Certificates
(Ο, V_Ο, R_Ο): Every row includes aprojection_certificatespecifying the projection operatorΟ, retained subspaceV_Ο, admissible revision classR_Ο, complementary re-grounding classR \ R_Ο(ccr://headroom_retrieve), and Monotone Kernel Growth Guard (ker Ο_n β ker(Ο_{n+1} β Ο_n)). - Content-Addressed P2P Swarm (
n0-computer/iroh): Every row is stamped with its BLAKE3 content-addressediroh_ticket(bloba...) andiroh_blake3_hash.
Schema (data/train.jsonl)
Each row contains the raw/compressed strings, MOEOptimizer routing tensor metadata, formal Revision-Sensitive Projection Certificate (Ο, V_Ο, R_Ο), Iroh BLAKE3 ticket, and a ready-to-train ShareGPT / OpenAI-format sft_messages array:
| Field | Type | Description |
|---|---|---|
id |
string |
Deterministic SHA-256 content hash |
timestamp |
string |
ISO-8601 UTC capture timestamp |
agent |
string |
Originating agent harness (opencode, crush, crush-dev-local, codewhale, claude, codex) |
tool_name |
string |
Tool or command identifier (e.g. bash:cargo_build, bash:pytest, mcp:query_production_events) |
transform |
string |
Compression layer (moe_optimizer:*, lean_context, smart_crusher, ccr) |
content_type |
string |
build_log, traceback, git_diff, json_array, markdown_table, source_code, tool_output |
raw_tokens |
int |
Token count before compression |
compressed_tokens |
int |
Token count after compression |
tokens_saved |
int |
Tokens removed while preserving all error/diagnostic signals |
savings_pct |
float |
Percentage token reduction (0-100) |
moe_primary_expert |
string |
Top-1 specialist expert selected by BurnMoeGatingRouter |
moe_secondary_expert |
string |
Top-2 specialist expert selected by BurnMoeGatingRouter |
moe_expert_weights |
dict |
Softmax routing distribution across all 6 experts |
moe_features |
dict |
7D structural density vector (json_density, tabular_density, error_signal_density, diff_patch_density, log_repetition_density, code_ast_density, prose_entropy) |
projection_certificate |
dict |
Formal (Ο, V_Ο, R_Ο) sufficiency certificate (projection_operator, retained_subspace, admissible_revisions, regrounding_required_for, monotone_kernel_guard, bennett_reversible_handle) |
iroh_ticket |
string |
n0-computer/iroh BLAKE3 content-addressed blob ticket (bloba...) |
iroh_blake3_hash |
string |
64-hex BLAKE3 digest of compressed payload |
raw_content |
string |
Sanitized original uncompressed tool output |
compressed_content |
string |
Sanitized compressed / signal-windowed tool output |
sft_messages |
list[dict] |
[system, user, assistant] chat messages for SFT / MLX / QLoRA fine-tuning |
Usage
from datasets import load_dataset
ds = load_dataset("PeetPedro/headroomz-compression-traces", split="train")
print(ds[0]["tool_name"], ds[0]["savings_pct"])
print(ds[0]["sft_messages"])
- Downloads last month
- 43