license: other
task_categories:
- text-generation
language: - en
tags: - calibration
- moe
- pruning
- reap
- glm-5.2
- agentic
- coding
size_categories: - 10K<n<100K
configs: - config_name: default
data_files: vera-keep.jsonl
splits:
train: 52694
pretty_name: Vera Agentic Coder
version: 1.1.0
Vera Agentic Coder - GLM-5.2 REAP-50 Calibration Corpus
Public-safe calibration corpus for a steep 50% REAP expert-pruning pass on GLM-5.2. The mix is weighted to preserve agent harness/tool behavior and coding capability, while retaining baseline general knowledge, basic math/reasoning, infra, security, shell, and home-lab competence.
This is intended for router/expert activation profiling only. It is not a supervised training set or benchmark.
Stats
| Records | 52,694 |
| Approx tokens | 10,974,534 |
| Max chars per row | 4,096 |
| Language/privacy filter | CJK removed; public-safety regex gate applied |
| SHA-256 | 99ddcceac999402f4b9bf9a4e725d98ca7b051b8d3557c63f83dde7e3d680b9a |
Token Mix By Role
| Role | Rows | Tokens | Share |
|---|---|---|---|
| agentic_traces | 6,363 | 4,088,477 | 37.3% |
| coding | 12,388 | 4,042,313 | 36.8% |
| baseline | 2,473 | 1,070,833 | 9.8% |
| math_reasoning | 18,050 | 891,265 | 8.1% |
| infra_security_shell | 13,171 | 826,383 | 7.5% |
| agent_harness | 249 | 55,263 | 0.5% |
Token Mix By Domain
| Domain | Rows | Tokens | Share |
|---|---|---|---|
| agentic | 6,479 | 4,123,887 | 37.6% |
| coding | 12,403 | 4,044,500 | 36.9% |
| general | 2,091 | 1,035,927 | 9.4% |
| math_reasoning | 18,050 | 891,265 | 8.1% |
| security | 734 | 637,451 | 5.8% |
| home-lab | 9,322 | 149,147 | 1.4% |
| business | 500 | 52,572 | 0.5% |
| shell | 3,115 | 39,785 | 0.4% |
Token Mix By Source
| Source | Rows | Tokens | Share |
|---|---|---|---|
| devops-qa | 5,686 | 1,761,191 | 16.0% |
| glaive-fc-v2 | 2,735 | 1,157,484 | 10.5% |
| hermes-agent-glm-5.1 | 1,130 | 1,148,378 | 10.5% |
| wiki-rewrite | 1,973 | 1,018,261 | 9.3% |
| basic-math-reasoning | 18,050 | 891,265 | 8.1% |
| magicoder-oss-instruct | 1,578 | 848,084 | 7.7% |
| evol-instruct-code | 1,532 | 776,716 | 7.1% |
| hermes-agent-kimi | 669 | 678,545 | 6.2% |
| clawhub-security | 734 | 637,451 | 5.8% |
| swe-bench-oracle | 822 | 611,575 | 5.6% |
| codefeedback-filtered | 801 | 419,535 | 3.8% |
| hermes-fc-v1 | 744 | 252,486 | 2.3% |
| agent-instruct | 263 | 240,009 | 2.2% |
| home-assistant-requests | 9,320 | 147,099 | 1.3% |
| codealpaca-20k | 2,043 | 145,496 | 1.3% |
| python-code-instructions | 748 | 91,291 | 0.8% |
| vera-eval-harness | 226 | 53,930 | 0.5% |
| lean-six-sigma | 500 | 52,572 | 0.5% |
| linux-commands | 3,030 | 38,857 | 0.4% |
| proxmox-docs | 2 | 2,048 | 0.0% |
| agent-harness-registry | 23 | 1,333 | 0.0% |
| macos-shell | 85 | 928 | 0.0% |
Format
{"id": "uuid", "text": "raw calibration text...", "domain": "agentic", "source": "hermes-agent-glm-5.1", "tokens_approx": 512, "mix_role": "agentic_traces"}
Public-Safety Filter
Before publication, rows were dropped if they matched high-risk regex classes: email address, phone-like number, SSN-like number, credit-card-like number, API/secret/password/token assignment, bearer token, private key block, local user path, IPv4 address, street-address-like string, or common placeholder full names.
See public-safe-filter-report.md for the filter summary and final distribution. This is a regex-based public-release gate, not a formal privacy audit or legal guarantee.
Intended Use
Feed vera-keep.jsonl as the KEEP-domain calibration set to record router/expert activations before REAP scoring. The dataset intentionally keeps agentic/coding dominant without removing basic math and general intelligence signals.
License
Mixed-source corpus. Individual records inherit upstream dataset licenses and usage constraints. Review upstream licenses before redistribution or commercial use.
Credits
- Curation: Andrew Hornsby / Vera
- Build execution: ModelLab