nx_eval_landscape_research.nx source
↩ module page · 30 lines · 2191 B
1// nx_eval_landscape_research.nx -- THIN research organ: what Unsloth + the standard eval tools RECOMMEND TESTING,
2// so our sovereign capability suite covers EVERYTHING they do (correctness-vs-reference, perplexity, the benchmark
3// battery: MMLU/HellaSwag/GSM8K/...). Composes nx_research_engine (shared fetch->extract->bank). No TSV, no copy-paste.
4// expect_exit: 0 license_tier: ORIGINAL
5import "nx_research_engine.nx"
6const K_MAGIC_8388608: i64 = 8388608
7
8func main() -> i64 {
9 let store: *TrustStore = rf_init()
10 if (store as i64) == 0 { rf_puts("eval_landscape: trust store load failed\n" as *u8); return 1 }
11 rf_puts("CA roots="); rf_putn(trust_store_count(store)); rf_puts(" -- LLM EVAL/TEST battery -> Library\n" as *u8)
12 let cap: i64 = K_MAGIC_8388608
13 let out: *u8 = sys_mmap(cap)
14 var ok: i64 = 0
15
16 rf_section("WHAT UNSLOTH + EVAL TOOLS RECOMMEND (correctness-vs-reference kernels, perplexity, throughput)" as *u8)
17 ok = ok + rf_fetch_bank("https://raw.githubusercontent.com/unslothai/unsloth/main/README.md" as *u8, "eval_unsloth_readme" as *u8, store, out, cap)
18 ok = ok + rf_fetch_bank("https://raw.githubusercontent.com/EleutherAI/lm-evaluation-harness/main/README.md" as *u8, "eval_lm_harness_readme" as *u8, store, out, cap)
19
20 rf_section("STANDARD CAPABILITY BENCHMARKS (knowledge / commonsense / reasoning / the metric)" as *u8)
21 ok = ok + rf_fetch_bank("https://en.wikipedia.org/wiki/MMLU" as *u8, "eval_mmlu" as *u8, store, out, cap)
22 ok = ok + rf_fetch_bank("https://en.wikipedia.org/wiki/HellaSwag" as *u8, "eval_hellaswag" as *u8, store, out, cap)
23 ok = ok + rf_fetch_bank("https://en.wikipedia.org/wiki/Perplexity" as *u8, "eval_perplexity" as *u8, store, out, cap)
24
25 rf_section("NON-NVIDIA ACCELERATOR BENCHMARK (hardware-agnostic reference: AWS Inferentia/Trainium via the Neuron SDK)" as *u8)
26 ok = ok + rf_fetch_bank("https://raw.githubusercontent.com/aws-neuron/aws-neuron-sdk/master/README.md" as *u8, "eval_aws_neuron" as *u8, store, out, cap)
27
28 rf_puts("EVAL-LANDSCAPE BANKED: "); rf_putn(ok); rf_puts(" / 6 (sections preserved, gzip-guarded; run nx_library_harvest_v2 to index)\n" as *u8)
29 return 0
30}