GLM-5.3-Flash-NVFP4-FP8ATTN-512K / scripts /results-longrep-recal.json
tacos4me's picture
model card (WIP banner), serving patches + SM120 tuned config, conversion/acceptance scripts, LICENSE
513fb85 verified
Raw
History Blame Contribute Delete
1.64 kB
{
"model": "glm-5.3-flash-recal",
"ctx": 260000,
"gen": 2200,
"rounds": [
{
"round": 1,
"prompt_tokens": 225032,
"completion_tokens": 2712,
"wall_s": 64.1,
"finish_reason": "length",
"four_gram_max": 2,
"four_gram_top": "at least 2,000 words",
"eight_gram_max": 1,
"eight_gram_top": "The user wants a detailed essay of at",
"max_consec_line_repeat": 1,
"degenerate_4gram_ge12": false,
"tail": "hmarks but degenerate during extended generation. The most informative evaluations combine these approaches, testing models under conditions that approximate real-world usage while maintaining the controlled variables needed for scientific comparison. As context windows continue to expand and models are deployed in increasingly demanding applications, the sophistication of evaluation methodologies"
},
{
"round": 2,
"prompt_tokens": 224997,
"completion_tokens": 2712,
"wall_s": 58.4,
"finish_reason": "length",
"four_gram_max": 3,
"four_gram_top": "- Why long context",
"eight_gram_max": 1,
"eight_gram_top": "The user has provided a long filler document",
"max_consec_line_repeat": 1,
"degenerate_4gram_ge12": false,
"tail": "asymmetry; per-channel/per-token grouping; outliers\n- Eviction/compression alternatives (sinks, heavy hitters) as complementary\n- Evaluation protocol: sweep precision, measure NIAH/perplexity deltas, latency, memory\n- Sensitivity of retrieval to key perturbations; softmax sensitivity\n- Practical guidance: keep K higher precision or per-channel scales; quantize V aggressively; validate at multiple\n"
}
],
"verdict": "PASS"
}