dancinlab/hexa-codex-sandbox-evals-v1

hexa-codex SANDBOX Benign Evals v1

Benign evaluation / benchmark input sets from the hexa-codex SANDBOX domain. This dataset bundles the load-generation inputs for the OPS SLO-under-load benchmarks and the SUBSTRATE persona scale-ladder eval set. All prompts here are benign (arithmetic, definitions, general knowledge, persona-conditioned Q&A). Adversarial refusal-evaluation sets are kept in a separate private repo and are NOT included here.

What this measures

• SLO under load (OPS). A fixed prompt manifest is replayed against a long-lived llama-server (continuous batching, -cb) at a sweep of offered arrival rates and parallel-slot counts to measure p50 / p95 / p99 / p999 latency, throughput saturation, the M/M/c queue knee, and the accuracy-collapse-under-saturation signature.
• Persona scale ladder (SUBSTRATE). A stratified sample of the same manifest is run through three difficulty personas (nano / mid / max — mock haiku / sonnet / opus tiers) across a 0.5B → 1.5B → 3B → 7B model ladder to measure per-tier accuracy and latency separation at scale.

Provenance

Produced by the hexa-codex SANDBOX bench harnesses (local llama-server / llama-completion, Qwen2.5 GGUF on an M3 Metal mac-mini, $0 local serving).

File	Produced by (bench .hexa)	Cycle	Verdict it scored
slo_load_manifest.jsonl	bench/sandbox_stage2_manifest_gen.hexa (consumed by sandbox_stage4_slo_under_load.hexa and sandbox_stage4_slo_full_grid.hexa)	cycle-15 / cycle-15c	.verdicts/sandbox/stage4_slo_under_load_summary.txt (M2.OPS)
slo_grid_cells.jsonl	bench/sandbox_stage4_slo_under_load.hexa (9 cells) + bench/sandbox_stage4_slo_full_grid.hexa (18 cells)	cycle-15 / cycle-15c	.verdicts/sandbox/stage4_slo_under_load_summary.txt (M2.OPS); the 18-cell full-grid (M3.OPS) writes .verdicts/sandbox/m3_ops_full_slo_grid.tsv
persona_ladder_cases.jsonl	bench/sandbox_stage2_persona_scaled{,_1_5b,_3b,_7b}.hexa	cycle-6 / cycle-8 / cycle-13	.verdicts/sandbox/stage2_persona_scaled_summary.txt, ..._1_5b_summary.txt, ..._3b_summary.txt, ..._7b_summary.txt (M3.SUBSTRATE)
persona_tiers.jsonl	persona definition block (verbatim, shared across the four persona benches)	—	(definition file; same verdicts as above)

The .verdicts/sandbox/* files referenced above live in the source repo (github.com/dancinlab/hexa-codex). They are the recompute-verified verdicts the SANDBOX governance links each claim to; they are NOT redistributed in this dataset (this repo ships the eval inputs only).

Files & row schemas

slo_load_manifest.jsonl — 2000 rows

The stratified prompt manifest replayed under load (word-count strata wc_5_15 … wc_101_200, 400 prompts each).

id                   string  e.g. "slo-1"
domain               string  "slo_load"
stratum              string  word-count stratum
word_count           int
prompt               string  the user prompt
expected             string  expected keyword for byte-exact-subset scoring
expected_difficulty  int     1..N difficulty tag

slo_grid_cells.jsonl — 27 rows

The offered-load grid: 9 under-load cells (np ∈ {1,2,4} × rate ∈ {5,20,100} qps) plus 18 full-grid cells (np ∈ {1,2,4} × rate ∈ {1,2,5,10,20,40} qps).

id               string
config           string  "under_load_9cell" | "full_grid_18cell"
bench            string  source bench .hexa
np               int     parallel slots (-np)
target_rate_qps  int     offered arrival rate
warmup_sec       int     30 (excluded from percentile rollup)
measure_sec      int     60 (included)
max_tokens       int     64
model            string  serving model

persona_ladder_cases.jsonl — 450 rows

One row per (persona × sampled prompt): 150 uniformly-stratified manifest prompts (stride 13, 30 per stratum × 5 strata) × 3 personas. The same 150-prompt input set was scored against the 0.5B / 1.5B / 3B / 7B model ladder.

id                   string  e.g. "persona-nano-m1"
domain               string  "persona_scale_ladder"
persona              string  "nano" | "mid" | "max"
system_prompt        string  persona system prompt
temperature          float   0.0
max_tokens           int     32 | 256 | 1024
mock_vendor_tier     string  "haiku" | "sonnet" | "opus"
manifest_idx         int     1-based index into slo_load_manifest
stratum              string
word_count           int
prompt               string
expected             string
expected_difficulty  int

persona_tiers.jsonl — 3 rows

The three-tier persona definition (the scale-ladder dimension).

persona           string
system_prompt     string
temperature       float
max_tokens        int
mock_vendor_tier  string

Usage

from datasets import load_dataset
slo = load_dataset("dancinlab/hexa-codex-sandbox-evals-v1", "slo_load_manifest", split="train")
persona = load_dataset("dancinlab/hexa-codex-sandbox-evals-v1", "persona_ladder_cases", split="train")

License

Apache-2.0. These are benign synthetic evaluation prompts intended for serving / latency benchmarking and small-model accuracy probing.