datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
wikitext-2wikitext-2-raw-v1-preprocessedwikitext_2_detokenizedwikitext-220728-250728wikitext-2-raw-v1qfs-smollm2-135m-wikitext2-campaign-v1
SmolLM2-135M QFS calibration and evaluation campaign
A small stored-weight fidelity study, not a broad model-quality benchmark.
Evaluation: 16 complete WikiText2 raw test articles, one 256-token window each, 4080 prediction positions.
Calibration: 32 disjoint train articles, 256 tokens each, 8192 calibration tokens.
Complete article title, normalized content and exact 13-token-ngram separation were checked. Validation is unused. Pretraining overlap remains unknown.
Original… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/qfs-smollm2-135m-wikitext2-campaign-v1.wikitext-2__llama__block-size-1024
Dataset Card for "wikitext-2__llama__block-size-1024"
More Information needed
qfs-smollm2-135m-wikitext2-native-v1
HF workflow d3dc69602aeb981f06bd9f4c726937f9
A root fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from malaiwah/SmolLM2-135M-QFS-native-bf16.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/qfs-smollm2-135m-wikitext2-native-v1.wikitext-2-split-128This is a dataset created from the WikiText-2 dataset by splitting longer sequences into sequences with maximum of 128 tokens after using a wordpiece tokenizer.
qfs-smollm2-135m-wikitext2-gptq-g32-v1
HF workflow 32c6ab05b0ceab1cecdceda838846388
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from malaiwah/SmolLM2-135M-QFS-gptq-int4-g32.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/qfs-smollm2-135m-wikitext2-gptq-g32-v1.qfs-smollm2-135m-wikitext2-gptq-g64-v1
HF workflow 73f0a12a901c7368794a3a886f55b675
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from malaiwah/SmolLM2-135M-QFS-gptq-int4-g64.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/qfs-smollm2-135m-wikitext2-gptq-g64-v1.wikitext-2-raw-v1-shuffled
Dataset Card for "wikitext-2-raw-v1-shuffled"
More Information needed
qfs-smollm2-135m-wikitext2-rtn-g64-v1
HF workflow ee13c03b2f256042a17c9817e17a2f20
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from malaiwah/SmolLM2-135M-QFS-rtn-int4-g64.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/qfs-smollm2-135m-wikitext2-rtn-g64-v1.wikitext-2-v1wikitext-2-ppl
WikiText-2 (raw) Perplexity Eval
Curated mirror of Salesforce/wikitext maintained by the majentik MLX-quantization project.
Why this exists (the plan it serves)
Standard perplexity benchmark for quant quality (the DWQ / perplexity gate reports PPL on this corpus to compare tiers and vs the BF16 baseline).
What this is
Full test split of wikitext-2-raw-v1. Column: text.
Upstream & license
Source: Salesforce/wikitext
License:… See the full description on the dataset page: https://huggingface.co/datasets/majentik/wikitext-2-ppl.wikitext-2-raw-v1-preprocessed-200-PI_KFI_claude-claude-3-5-sonnet-20240620wikitext-2-raw-v1-preprocessed-5k-PI_KFI_claude-claude-3-5-sonnet-20240620wikitext-2-raw-v1-preprocessed-1k-PI_KFI_claude-claude-3-5-sonnet-20240620Cloze_QA_Dataset_Wikitext2
Cloze QA Dataset (WikiText-2)
Dataset Description
The Cloze QA Dataset is automatically generated from the WikiText-2 corpus. It contains fill-in-the-blank (cloze) style questions derived directly from sentences in Wikipedia articles. This dataset is particularly useful for evaluating local recall, reading comprehension, and contextual understanding.
Each document produces exactly three unique QA pairs, preserving document structure and sentence alignment while… See the full description on the dataset page: https://huggingface.co/datasets/CGU-Widelab/Cloze_QA_Dataset_Wikitext2.wikitext-2-raw-v1wikitext2raw_benchmark_royaVALUE_wikitext2_been_done
Dataset Card for "VALUE_wikitext2_been_done"
More Information needed
wikitext-2-second-halfwikitext-2-raw-v1-merged-45kwikitext2_benchmarkVALUE_wikitext2_got
Dataset Card for "VALUE_wikitext2_got"
More Information needed
wikitext-2-raw-v1-test
wikitext-2-raw-v1-test
Test split for lm-eval-harness.
wikitext-2-raw-v1-64-token-blockswikitext2-raw-MIA_paraphrased_defensewikitext-2-raw-v1-forbidden-titles