thanhnx12/mla-moe-dataset-public

mla-moe public evaluation dataset

Public (participant-facing) input set for the MLA-MoE inference throughput exercise. Two target models, 512 prompts each, with the model's own greedy continuation as the correctness reference:

subset	target model
glm47/	zai-org/GLM-4.7-Flash
dsv2lite/	deepseek-ai/DeepSeek-V2-Lite

Grading is done on a held-out private set with the same size and length distribution, generated the same way but from disjoint source text. Tune against the public set; do not tune against a specific prompt.

Files (per subset)

file	content
requests.txt	the input. Line 1 = request count (512), then one prompt per line, verbatim single-line text.
prompts.i32.txt	prompt token ids, one line per prompt, space-separated (add_special_tokens=False).
completions.i32.txt	reference output — the model's greedy continuation, token ids, one line per prompt, ≤ 64 tokens.
reference.json	per-request {prompt_len, completion_len, hf_nll, hf_ntok} plus provenance (model_dir, max_new, dtype, transformers_version).
manifest.json	prompt-generation parameters: seed, length buckets, source book ids, length statistics, per-prompt target lengths.

Token ids are produced by each model's own tokenizer, so glm47/ and dsv2lite/ ids are not interchangeable — the prompt text differs too (independent length trimming per tokenizer).

Length characteristics

Prompt token lengths: min 64, median 256, max 512 (mean 241), drawn from buckets 64 / 128 / 256 / 384 / 512. Longest prompt + completion is 576 tokens, well inside the 4096-token budget the harness assumes.

How the reference was produced

Greedy decoding of the target model itself — HF transformers 5.12.1, fp32, max_new_tokens=64, no sampling. hf_nll is the teacher-forced negative log-likelihood of the reference continuation, hf_ntok the number of scored tokens.

The reference was regenerated at fp32 (it was previously bf16). The reason is that greedy decoding is a hard argmax, so any small numeric difference (dtype, kernel, quantization) can make a completion diverge from some token onward — and a bf16 reference diverged from an fp32-faithful engine on most prompts. At fp32 the reference is reproducible token-for-token by an fp32 engine, which makes it usable as a strict reference as well as a fuzzy one.

Concretely, against the previous bf16 reference an independent fp32 CPU engine matched only 3/8 sampled completions exactly and scored 96.6% teacher-forced top-1; against this fp32 reference the same engine matches 8/8 exactly and scores 100% top-1. Note that this also means an engine using bf16/fp8 weights or KV cache will still diverge from this reference on many prompts — score those with fuzzy metrics (METEOR + BERTScore-F1) rather than exact match.

Prompt provenance

Prompts are sampled from 12 public-domain Project Gutenberg books, cleaned into single-line paragraphs, then trimmed to a target token length with the model's own tokenizer. Generation is deterministic (seed recorded in manifest.json). Source book ids: 76, 84, 98, 174, 768, 1232, 1260, 1342, 1400, 1661, 2554, 2701.

The prompt text is public domain. The reference completions are outputs of GLM-4.7-Flash and DeepSeek-V2-Lite and are subject to those models' own licenses.

Loading

from huggingface_hub import snapshot_download

path = snapshot_download("thanhnx12/mla-moe-dataset-public", repo_type="dataset")

with open(f"{path}/glm47/requests.txt") as f:
    n = int(f.readline())
    prompts = [f.readline().rstrip("\n") for _ in range(n)]

with open(f"{path}/glm47/completions.i32.txt") as f:
    reference = [[int(t) for t in line.split()] for line in f]

assert len(prompts) == len(reference) == 512