MARS omission detection: exact research checkpoints

Frozen artifacts for the primary detector, its judge-label control, the MARS-2 research baseline and the distilled inventory segmenter. Seeds: 20260906, 20260907 and 20260908. Load the chosen checkpoint subdirectory.

Component Directory Scoring code
MARS-C (three seeds) results/marsc/e1/mc-A-lam1-crossenc-inv_seed*/model mc_score_units.py applies the original one-logit verifier convention.
Judge-label control (B21, three seeds) results/mars2_gates/b21/xenc_seed* m31_score_pipeline_all.py implements the original hypothesis construction and three-way scoring.
MARS-2 (six checkpoints) models/plan2026/b2/{direct_enabled,full_recovery}_seed*/epoch1 mars_v2.dump loads the backbone and the auxiliary heads.
LongT5 results/mars3/m32/segmenter replay/e5_segmenter.py implements historical preprocessing and decoding.

Verify and run

Download with the Hugging Face CLI and pin the commit with --revision for repeatability. From the snapshot root verify the downloaded weights:

sha256sum -c CHECKPOINTS.sha256

Smoke test on an illustrative example:

mkdir -p example-output
printf '%s\n' '{"pair_id":"example","candidate":"The ferry sails on Monday.","source":"The ferry sails on Monday and costs five dollars.","units":[{"text":"The fare is five dollars."}]}' > example.jsonl
python research/marsc_20260916/code/mc_score_units.py --units example=example.jsonl --models results/marsc/e1/mc-A-lam1-crossenc-inv_seed20260906/model --tag humanfact --scores-out example-output

The example writes example-output/example_humanfact_seed20260906.jsonl. In each row, scores follows the input unit order. A larger score indicates stronger evidence that the candidate omits the unit. In the ferry example, the scored unit is the fare. With the example input and seed above, the recorded omission score was approximately 0.997 (one value in the scores array).

For TEST, supply the two original derived inventories as gemma+ent=PATH and distilled=PATH after --units. Quote the --models glob to run all three seeds. For B21 use research/mars3_20260916/code/m31_score_pipeline_all.py with the same inventories and --ckpts 'results/mars2_gates/b21/xenc_seed*'. reference_scores/ contains the twelve original TEST scores as identifiers and numeric predictions. EXPECTED_OUTPUTS.sha256 lists these twelve outputs and six MARS-2 outputs. Compare filenames from the output directory. environment/ records the historical software and platform.

For MARS-2 use:

MARS2_SPACY_MODEL=en_core_web_sm python -m mars_v2.dump --model CHECKPOINT --pairs PAIRS --out OUTPUT --id-key pair_id --source-key source --candidate-key candidate --units entity+proposition

For segmenter comparison run replay/e5_segmenter.py ORIGINAL NEW_OUTPUT on CUDA. ORIGINAL contains segmenter/, eight props_shard files and the historical distilled_props.jsonl comparator. For inventory construction use research/mars3_20260916/code/m32_inventory.py with B1_DIR pointing to the split's pairs directory.

Benchmark inputs

Obtain RoSE through the original repository. Follow the underlying CNN/DailyMail, XSum and SAMSum terms for source text. INPUTS.sha256 identifies the derived files used by the historical run. The benchmark download alone does not contain frozen MARS proposition shards or inventories. Byte-level reproduction requires matching inputs. The MARS evaluation package includes the per-table map. This model release omits original benchmark text and individual annotator responses.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Glazkov/MARS-omission-checkpoints

Finetuned
(46)
this model