Datasets:

---

hotchpotch

/

s1mb-result

Like 1

Leaderboard comparison

Scores: baseline-adjusted 0–100, higher is better; equal benchmark weights. Optional Box/gen-box use mean IoU ×100 and never affect ranking. Noul uses balanced accuracy. Gen is the Diverse/Contextual subset of each task. Static Active Params excludes embedding weights, including tied output weights (embedding_excluded_parameters_v1); historic results retain their recorded method. N/A means unknown parameters or unavailable scores.

All 1 rows: 137/137 complete, zero failed decisions (required non-Box scope; Box optional).
Validated against each result's recorded dataset revision.

Submission evidence

Evaluation category: english-v1.

Leaderboard submission requires every active benchmark in the selected scope (Box is optional and excluded from ranking) complete and measured, across Choice, Noul and Score, with zero failed decisions. Missing benchmarks, unfinished tasks and partial results are not leaderboard-eligible. Validation success alone does not establish complete coverage. For input, capacity or API failures, follow the adapter failure guide and complete a corrected evaluation before submitting.

PelaAI__KnowLine-0.8B-Gen1

Recorded provenance and effective settings (distinct configurations, with per-benchmark API usage counts left out; they are in the result files):
Per-example hashes and image receipts remain in the result files; this report summarizes image policies.

[
  {
    "run_id": "knowline-08b-gen1-en-20261029-001",
    "model": {
      "training_exposure": null,
      "id": "PelaAI/KnowLine-0.8B-Gen1",
      "adapter": "knowline",
      "revision": "b76c79d4c64d35219b7801a66a5f84cc718dce4e",
      "settings": {
        "requested_model": "PelaAI/KnowLine-0.8B-Gen1",
        "degenerate_box_policy": "retain-zero-area-score-zero-iou",
        "vision_transport": "multipart-full-image-v1",
        "score_criteria_order": "ascending-value-v1",
        "renderer": "system-one-structured-v2",
        "input_length_policy": "full-input-no-local-truncation",
        "probability_normalization": "server softmax over option labels; sum-normalized within 1e-6",
        "server": "published knowline_server.py behind SGLang 0.5.21, bf16, chat style, temperature 1",
        "served_model_name": "PelaAI/KnowLine-0.8B-Gen1",
        "concurrent_cases": 16,
        "questions_per_request": 64,
        "weight_dtype": "bfloat16",
        "quantization": null,
        "read_twice": true,
        "twice_max_tokens": 3000,
        "folded_readout_temperature": 1.54,
        "per_type_temperatures": {
          "score": 0.6494
        },
        "server_mem_fraction_static": 0.85,
        "mamba_radix_cache_strategy": "extra_buffer",
        "enable_fp32_lm_head": true,
        "front_end_workers": 16,
        "release_file_sha256": {
          "knowline_server.py": "e1dfccb3f15dcdea439dc76bb7fe055202199fecf0db6bcfc4aca725e694bd3a",
          "knowline_layout.py": "2e839fa24f1ad9c980898b2a9aa62e2154ba3f57618bd5f2fedb5756f61f062f",
          "serve_knowline.sh": "79c4568a3f7ae97fb64ff949aa478ade36ca68c4beacf4bcc3056f38aa696fe4",
          "temperatures.json": "aef62724291c261df3b42c6f51d22b11abe6515bd87adf077d55a042d5fb0283"
        },
        "metadata_correction": "Only replace upstream hard-coded FP8 description; prediction/scoring code unchanged",
        "metadata_wrapper_sha256": "d67d4012ae9fd0e0236767d36ff59949c8253ef7b7886bec731a269312e75c43"
      },
      "total_params": null,
      "active_params": null,
      "parameter_count_method": null
    },
    "evaluator_version": "0.1.0",
    "provenance": "measured",
    "environment": {
      "python": "3.11.16",
      "platform": "Linux-5.4.119-19-0009.56-x86_64-with-glibc2.35",
      "evaluator_source_sha256": "27ecceb6bd177d7f089c0a410cf734df81c838d9976094e5339ad9d8590fb1fe",
      "evaluator_revision": "9ad06feacc372af4a99beea0bf25a0bcaa1df5c4",
      "dataset_source": {
        "repo_id": "hotchpotch/S1MB-english-v1",
        "revision": "10bcb8c6916b98b966cc1f03aab1781a7c354b4c"
      }
    }
  }
]

Training source and domain disclosure

For English v1 and Japanese v1, please disclose, to the best of your knowledge, whether the submitted model was trained on any original evaluation source family, directly or through mirrors, derivatives or mixtures, as well as the Bekko System One Dataset, the Saihai Dataset, or other datasets covering the same tasks or source domains as these evaluations.

Review the English/Japanese original dataset and source checklist, which maps active benchmark IDs to HF dataset names and sources outside HF. English examples include PolyAI/banking77, clinc/clinc_oos, fancyzhx/dbpedia_14, Anthropic/hh-rlhf and LocalLLaMA/typed-decisions; Japanese examples include JGLUE, JSICK, sbintuitions/JamC-QA, openai/MMMLU and hakari-bench/NanoJMTEB-v2. These examples do not replace the full checklist. Not using Bekko System One or Saihai does not establish separation from S1MB source families or evaluation inputs.

Identify the source families actually used, the affected S1MB benchmarks, and the training revisions, configurations, splits and filtering when known. Link model cards, training manifests, code or technical reports supporting the disclosure. Record supported benchmark-level declarations in model metadata as training_exposure, distinguishing same_source_family, related_domain and evaluation_examples. Reconcile audit findings with metadata and prose: distinguish confirmed evaluation-input use from approximate-match candidates, and do not claim no evaluation-input use when confirmed matches are reported. Training on an NLI dataset, for example, does not by itself establish exposure to every NLI benchmark. Filtering evaluation duplicates does not remove source-family exposure.

If source details cannot be confirmed, state what is unknown. Missing disclosure does not mean a domain was unseen: other models may also have trained on these domains without public confirmation. Known-domain marks communicate documented training exposure, not an explanation or prediction of high scores; scores are unchanged.

Author explanation and assessment

Author explanation

This adds the complete English S1MB measurements for KnowLine-0.8B-Gen1, a Qwen3.5-0.8B LoRA decision model trained for one epoch on mixU (the training data of KnowLine-9B-Gen1) and merged into bf16 weights; the released checkpoint is the end of training. Readout temperature T=1.54 is folded into the norm. No leaderboard-first claim is made.

Coverage and fixed settings

• All 137/137 English-v1 benchmarks complete: 18,701 cases, 26,269 successful decisions, zero failed and zero abstained. All 57 Choice, 59 Noul and 21 Score benchmarks are present. This is English text scope only, not Japanese or vision.
• Measured model and inference-code revision: b76c79d4c64d35219b7801a66a5f84cc718dce4e. The later documentation commit 6b27831f7b783c03278bd946d4d8f4f7a98a3798 (model card, inference notes, TRAINING_SOURCES.json, checksums) does not change the evaluated weights or serving code.
• Dataset: hotchpotch/S1MB-english-v1 @ 10bcb8c6916b98b966cc1f03aab1781a7c354b4c.
• Evaluator: 9ad06feacc372af4a99beea0bf25a0bcaa1df5c4. Raw run: knowline-08b-gen1-en-20261029-001.
• One H20 GPU, released native /v1/systemone front end behind bf16 SGLang 0.5.21, Transformers 5.12.1, read twice for text-only requests up to 3,000 tokens, thinking off, temperature 1 after folded T=1.54. The released temperatures.json divides score-question logits by 0.6494 (= 1/1.54), so score questions are read at an effective temperature of 1.0; choice and yes/no questions keep the folded temperature. No Header. Sixteen front-end workers and sixteen concurrent cases. Overflow is rejected, not truncated; all supplied options are retained.
• The standard KnowLine adapter preserves dataset-default structured instructions and ascending numeric Score levels. The installed adapter's metadata hard-codes an FP8 server description. A metadata-only wrapper records the actual bf16 / read-twice / per-type-temperature configuration and source-file hashes; inference, normalization and official scoring are unchanged. That wrapper's hash is recorded in each result.
• Temperature, checkpoint, prompts and serving settings were fixed before this S1MB test. No test-based tuning, rescoring to choose a temperature, or per-benchmark selection was performed.
• A bounded smoke covered two cases for each task; its results validate with zero errors. The full run was not overwritten or selectively patched.
• Official raw-run validation: 137 files / 137 unique results. Official export validation: 137 published results.

Training exposure

metadata.json declares supported exposure for 71 distinct benchmarks, the same declaration as KnowLine-9B-Gen1 because the training data is the same. The evidence is a pinned retained-channel/source-family manifest, not scores. mixU contains 1,143,656 rendered rows across 720 normalized channels: retained earlier decision data, previously unseen rows of the same sources, injection-invariance examples and paired-convention pilot rows.

Public benchmark source-family train data and related-domain synthetic data are disclosed even after decontamination. The actual retained training corpus is not public. Historical upstream revisions and some inherited split details remain unknown. Mixed-channel totals are not represented as exact per-source counts. We do not copy another model's record-level overlap counts or claim a new independent record-level S1MB overlap audit. Unlisted exposure and base-model pretraining are unknown, not asserted absent. No earlier model's declaration is modified by this submission.

Results and interpretation

Task Avg is 34.1666, using the official equal-task mean. The six Diverse/Contextual benchmarks have a Task Avg of 66.3723; this is a small defined subset and not evidence of broad unseen-task performance. No Borda score or rank is claimed because it depends on the comparison cohort.

The model weights follow the published model license. Submitted measurements and their source-derived contents retain applicable evaluation dataset/source terms; this submission does not redistribute benchmark inputs or claim a new license for upstream datasets.

Thank you @PEScn for contributing the KnowLine-0.8B-Gen1 results. The reviewed submission passes validation and the mandatory full English completion check (137/137 benchmarks, zero failed decisions), and its pinned training-source declarations are retained as known-domain marks. Accepted and merged; the leaderboard deployment follows.