Vveerrgg's picture
v1
ed9f38e verified
|
Raw History Blame Contribute Delete
13.7 kB
metadata
license: apache-2.0
language:
  - en
  - fr
pretty_name: Qwen3.8-27B Bilingual Canadian evaluation set
tags:
  - evaluation
  - gguf
  - llama.cpp
  - french
  - quebec
configs:
  - config_name: v1
    data_files:
      - split: belebele_gen_eng_Latn
        path: v1/tasks/belebele_gen_eng_Latn.jsonl
      - split: belebele_gen_fra_Latn
        path: v1/tasks/belebele_gen_fra_Latn.jsonl
      - split: ifeval
        path: v1/tasks/ifeval.jsonl
      - split: ifeval_fr
        path: v1/tasks/ifeval_fr.jsonl
      - split: gsm8k
        path: v1/tasks/gsm8k.jsonl
      - split: mgsm_hj_en
        path: v1/tasks/mgsm_hj_en.jsonl
      - split: mgsm_hj_fr
        path: v1/tasks/mgsm_hj_fr.jsonl
      - split: cole_allocine
        path: v1/tasks/cole_allocine.jsonl
      - split: cole_fquad
        path: v1/tasks/cole_fquad.jsonl
      - split: cole_fracas
        path: v1/tasks/cole_fracas.jsonl
      - split: cole_gqnli
        path: v1/tasks/cole_gqnli.jsonl
      - split: cole_mms
        path: v1/tasks/cole_mms.jsonl
      - split: cole_multiblimp
        path: v1/tasks/cole_multiblimp.jsonl
      - split: cole_paws_x
        path: v1/tasks/cole_paws_x.jsonl
      - split: cole_piaf
        path: v1/tasks/cole_piaf.jsonl
      - split: cole_qfrblimp
        path: v1/tasks/cole_qfrblimp.jsonl
      - split: cole_qfrcola
        path: v1/tasks/cole_qfrcola.jsonl
      - split: cole_rte3
        path: v1/tasks/cole_rte3.jsonl
      - split: cole_sickfr
        path: v1/tasks/cole_sickfr.jsonl
      - split: cole_xnli
        path: v1/tasks/cole_xnli.jsonl
source_licenses:
  belebele_gen_eng_Latn: >-
    CC-BY-SA-4.0; https://huggingface.co/datasets/facebook/belebele; text
    included: yes
  belebele_gen_fra_Latn: >-
    CC-BY-SA-4.0; https://huggingface.co/datasets/facebook/belebele; text
    included: yes
  ifeval: >-
    Apache-2.0; https://huggingface.co/datasets/google/IFEval; text included:
    yes
  ifeval_fr: >-
    Apache-2.0; https://huggingface.co/datasets/jzhang86/fr_ifeval; text
    included: yes
  gsm8k: 'MIT; https://huggingface.co/datasets/openai/gsm8k; text included: yes'
  mgsm_hj_en: >-
    CC-BY-SA-4.0; https://huggingface.co/datasets/juletxara/mgsm; text included:
    yes
  mgsm_hj_fr: >-
    CC-BY-SA-4.0; https://huggingface.co/datasets/juletxara/mgsm; text included:
    yes
  cole_allocine: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_fquad: >-
    CC-BY-NC-SA-3.0 (FQuAD);
    https://huggingface.co/datasets/graalul/COLE-public; text included: no
  cole_fracas: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_gqnli: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_mms: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_multiblimp: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_paws_x: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_piaf: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_qfrblimp: >-
    CC-BY-NC-4.0 (QFrBLiMP, evaluation use);
    https://huggingface.co/datasets/graalul/COLE-public; text included: no
  cole_qfrcola: >-
    CC-BY-NC-SA-4.0 (QFrCoLA, evaluation use);
    https://huggingface.co/datasets/graalul/COLE-public; text included: no
  cole_rte3: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_sickfr: >-
    upstream terms unclear; https://huggingface.co/datasets/graalul/COLE-public;
    text included: no
  cole_xnli: >-
    CC-BY-NC-4.0 (XNLI); https://huggingface.co/datasets/graalul/COLE-public;
    text included: no

Qwen3.8-27B Bilingual Canadian: evaluation set

This is the evidence behind every number published for Qwen3.8-27B Bilingual Canadian, a small GGUF file of Qwen3.8-27B for English and French. Each version of the model gets its own config; this is config v1.

For every benchmark item you get the item id, each model's answer and its score. Where the benchmark's licence allows it, you also get the prompt exactly as the model saw it and the reference answer. On top of that:

  • v1/scores.json: the score per task per model, recomputed from the per-item rows.
  • v1/kld.json: how close each file's next-token probabilities are to the original's (KL divergence and perplexity), per language.
  • v1/code/: the task configs and the MGSM answer extraction, so you can rerun or rescore everything.

The three models

Column prefix Model
original_bf16 Qwen3.8-27B, original BF16 weights (the reference)
this_build Qwen3.8-27B Bilingual Canadian v1 (GGUF, 9370.1 MiB)
unsloth_ud_q2_k_xl Unsloth's Qwen3.8-27B UD-Q2_K_XL (GGUF, 9373.6 MiB), the closest file in size

How the scores were produced

  • Every model ran in stock llama.cpp llama-server on one GPU, behind its OpenAI-compatible chat endpoint.
  • lm-evaluation-harness drove it with the local-chat-completions model type: thinking off, temperature 0, no sampling.
  • Letter-choice tasks (Belebele and most COLE tasks) score the first standalone letter in the reply.
  • GSM8K is the stock 5-shot task; the headline is the flexible-extract match.
  • MGSM is zero-shot with an answer extraction that reads English and French number formats (v1/code/mgsm_extract.py). The stock MGSM regex wants "The answer is N", which none of the three files writes in chat mode.
  • IFEval and French IFEval use the stock IFEval scorer; the headline is prompt-level strict accuracy.
  • FQuAD and PIAF are extractive question answering, scored by F1 against the gold spans.
  • KL divergence: llama-perplexity --kl-divergence against the original BF16 logits, 64 chunks of 1,024 tokens per language. "Same top" is how often the most likely next token matches the original's.
  • COLE tasks use the dev/validation splits of COLE-public, so they are not comparable to the COLE leaderboard, which scores hidden test splits. Big splits were cut to their first 2,000 rows (allocine, mms); FQuAD and PIAF keep rows that have a gold answer; FraCaS keeps rows whose answer is yes, no or unknown.

Files and columns

Each v1/tasks/<task>.jsonl row has task, item_id (lm-eval doc_id: the row's position in the split after the filtering above), doc_hash (lm-eval's SHA-256 of the source row, to match it exactly), and for every model <model>__answer (the raw reply), <model>__extracted (what the scorer read) and <model>__score (per-item metrics). Where text is included there are also prompt (chat messages as sent) and reference. IFEval rows add source_key (the dataset's own key); their reference is the instruction list and arguments the scorer checks.

Results (headline metric)

Task Original BF16 This build Unsloth
Belebele eng_Latn 0.962 0.961 0.959
Belebele fra_Latn 0.949 0.947 0.936
IFEval EN, prompt-level strict 0.821 0.806 0.830
IFEval FR, prompt-level strict 0.738 0.726 0.732
GSM8K, 5-shot, flexible match 0.964 0.955 0.953
MGSM EN, zero-shot 0.956 0.964 0.956
MGSM FR, zero-shot 0.884 0.892 0.888
COLE allocine 0.972 0.973 0.975
COLE fquad (F1) 0.524 0.557 0.618
COLE fracas 0.805 0.798 0.824
COLE gqnli 0.519 0.519 0.370
COLE mms 0.747 0.742 0.743
COLE multiblimp (not in EVALS.md) 1.000 1.000 1.000
COLE paws_x 0.808 0.791 0.792
COLE piaf (F1) 0.402 0.426 0.464
COLE qfrblimp 0.895 0.839 0.879
COLE qfrcola 0.800 0.761 0.756
COLE rte3 0.865 0.851 0.825
COLE sickfr 0.834 0.867 0.842
COLE xnli 0.762 0.713 0.745

GSM8K empty replies (the model ended its turn at once): This build 2, Original BF16 0, Unsloth UD-Q2_K_XL 0 of 1,319.

COLE multiblimp is published here but is not in EVALS.md, so its numbers were not cross-checked.

Closeness to the original (from v1/kld.json)

Text Model PPL Mean KLD 99th pct KLD Same top
French: FineWeb-2 fra_Latn, held-out test split Original BF16 6.03 0.0000 0.00 100.0%
French: FineWeb-2 fra_Latn, held-out test split This build 6.39 0.0928 0.85 86.5%
French: FineWeb-2 fra_Latn, held-out test split Unsloth UD-Q2_K_XL 6.42 0.0935 0.78 86.3%
Quebec French: FineWeb-2 documents from Quebec domains, held out Original BF16 6.70 0.0000 0.00 100.0%
Quebec French: FineWeb-2 documents from Quebec domains, held out This build 7.12 0.0908 0.78 86.4%
Quebec French: FineWeb-2 documents from Quebec domains, held out Unsloth UD-Q2_K_XL 7.13 0.0913 0.73 85.3%
English: held-out FineWeb-Edu web text Original BF16 7.69 0.0000 0.00 100.0%
English: held-out FineWeb-Edu web text This build 8.18 0.0908 0.85 87.0%
English: held-out FineWeb-Edu web text Unsloth UD-Q2_K_XL 8.16 0.0790 0.72 87.4%

Benchmark licences

Our own content here (scores, answers' packaging, card, code) is Apache-2.0. Benchmark text keeps its source licence. Text is included only where that licence plainly allows redistribution, with attribution below. Where the terms are non-commercial, evaluation-only or unclear, this set carries only item ids, doc_hash, model answers and scores; fetch the original rows from the source and match them by item_id (after the filtering described above) or doc_hash. Note that for extractive QA (FQuAD, PIAF) a model answer may itself quote a few words of the source passage.

Task Source Licence Question text included How to fetch
belebele_gen_eng_Latn Belebele (eng_Latn) CC-BY-SA-4.0 yes facebook/belebele, config eng_Latn, split test
belebele_gen_fra_Latn Belebele (fra_Latn) CC-BY-SA-4.0 yes facebook/belebele, config fra_Latn, split test
ifeval IFEval Apache-2.0 yes google/IFEval, split train
ifeval_fr IFEval French Apache-2.0 yes jzhang86/fr_ifeval, split train
gsm8k GSM8K MIT yes openai/gsm8k, config main, split test (5-shot from train)
mgsm_hj_en MGSM (en) CC-BY-SA-4.0 yes juletxara/mgsm, config en, split test
mgsm_hj_fr MGSM (fr) CC-BY-SA-4.0 yes juletxara/mgsm, config fr, split test
cole_allocine COLE allocine upstream terms unclear no allocine/allocine_validation.jsonl, first 2000 rows
cole_fquad COLE FQuAD CC-BY-NC-SA-3.0 (FQuAD) no fquad/fquad_validation.jsonl, rows with a gold answer
cole_fracas COLE FraCaS upstream terms unclear no fracas/fracas_test.jsonl, rows whose answer is yes/no/unknown
cole_gqnli COLE GQNLI-fr upstream terms unclear no gqnli/gqnli_validation.jsonl
cole_mms COLE MMS upstream terms unclear no mms/mms_dev.jsonl, first 2000 rows
cole_multiblimp COLE MultiBLiMP upstream terms unclear no multiblimp/multiblimp_dev.jsonl
cole_paws_x COLE PAWS-X upstream terms unclear no paws_x (dev file in COLE-public)
cole_piaf COLE PIAF upstream terms unclear no piaf/validation_piaf.jsonl, rows with a gold answer
cole_qfrblimp COLE QFrBLiMP CC-BY-NC-4.0 (QFrBLiMP, evaluation use) no qfrblimp/qfrblimp_dev.jsonl
cole_qfrcola COLE QFrCoLA CC-BY-NC-SA-4.0 (QFrCoLA, evaluation use) no qfrcola/frcola_dev.jsonl
cole_rte3 COLE RTE3-fr upstream terms unclear no rte3-french/rte3-french_dev.jsonl
cole_sickfr COLE SICK-fr upstream terms unclear no sickfr/sick_fr_validation.jsonl
cole_xnli COLE XNLI (fr) CC-BY-NC-4.0 (XNLI) no xnli/xnli_fr_validation.jsonl

COLE: COLE-public declares MIT for the compilation, but each subset comes from an upstream corpus with its own terms, which COLE does not restate; not redistributed here. Rows with "upstream terms unclear" are withheld for that reason. QFrBLiMP (CC-BY-NC-4.0) and QFrCoLA (CC-BY-NC-SA-4.0) are research and evaluation resources; XNLI is CC-BY-NC-4.0 and FQuAD CC-BY-NC-SA-3.0, so those are withheld as non-commercial. COLE subsets are fetched from graalul/COLE-public with the file named above.

Attribution for included text: Belebele (Bandarkar et al., 2023, Meta, CC-BY-SA-4.0); IFEval (Zhou et al., 2023, Google, Apache-2.0); French IFEval (jzhang86/fr_ifeval, Apache-2.0); GSM8K (Cobbe et al., 2021, OpenAI, MIT); MGSM (Shi et al., 2022, CC-BY-SA-4.0, built on GSM8K). The CC-BY-SA-4.0 text (Belebele, MGSM rows) stays under CC-BY-SA-4.0 in this set.

What is not here

Anything about how the model file was made. This set only measures the finished files.