SASLM-Eval

A 466-item benchmark testing whether a language model has learned the ontology of Sri Aurobindo's philosophical writing, as opposed to its surface style.

Perplexity measures fluency. This benchmark was built because perplexity on this corpus turned out not to be informative about conceptual accuracy, and there was no other way to measure it.

Scoring is by model likelihood only. No API, no judge model, no generation. The same checkpoint always produces the same number.

Tasks

Cloze (406 items, chance 25%)

A passage from held-out text with one ontological term removed, against three distractors drawn from the same semantic family. Selecting correctly requires knowing which member of the family belongs, not that a noun belongs.

Family	Items	Members
sanskrit_metaphysics	151	Brahman, Purusha, Prakriti, Sachchidananda, Maya, Jiva, Atman, Ananda
states	73	Ignorance, Knowledge, Bliss, Nescience
parts_of_being	72	psychic being, vital being, mental being, physical being
planes	56	Supermind, Overmind, Inconscient, Superconscient, Subconscient, Supramental
powers	54	Shakti, Ishwara, Divine Mother, Purushottama

Example (abbreviated): "…a general description of the passage from Overmind to ___" — options: Subconscient, Supramental, Superconscient, Supermind.

Style (60 items, chance 50%)

A real held-out sentence against a random permutation of its own words.

This task is saturated. Every model tested, including zero-shot GPT-2, scores 100%. It is retained for completeness but measures nothing. Do not report it as a result.

Scoring

Three rules are provided. Use PMI for cloze.

Rule	Definition
sum	Total log P(option | context). Biased toward short options.
byte	Divided by UTF-8 byte length (lm-eval-harness acc_norm).
pmi	log P(option | context) − log P(option | null context).

Mean per-token log-probability is deliberately not offered. It favours rare multi-token options, because the later subwords of a rare word are nearly determined once the first is emitted. Under that rule every model we tested scored at chance and never selected common terms — it measures tokenization, not knowledge. This is surface-form competition (Holtzman et al., 2021), and PMI is the standard correction.

For the style task PMI is undefined in practice, since there is no context to condition on. Use sum there.

Usage

python run_saslm_eval.py --checkpoint path/to/model --json results.json

The scorer accepts either a local checkpoint or a HuggingFace model id, and scores each model in its own vocabulary. Accuracy is a rank over the same options, so it remains comparable across different tokenizers — unlike perplexity.

Reported results

Cloze, PMI, 406 items, chance 25%, binomial margin ±4.7:

Model	Accuracy
SASLM-18M, primary only (3 seeds)	36.0 / 29.1 / 35.5
SASLM-18M, +20% commentary (3 seeds)	34.2 / 29.8 / 28.8
SASLM-18M, commentary only	17.7
GPT-2-small, zero-shot	30.3
GPT-2-small, fine-tuned	33.7

Read this carefully. Every model except the commentary-only one is above chance, so the benchmark does detect something. But no model is distinguishable from another. The between-seed range within a single condition is 6.9 points, larger than the binomial margin, so differences under about seven points carry no information. The ordering also changes under the other scoring rules.

The commentary-only model scores 17.7%, significantly below chance. A model trained only on commentary is worse than random at picking the term the primary author used.

Known limitations

• Underpowered for model comparison. See above. It can tell you a model is above chance; it cannot rank two models that are close.
• The style task is saturated and should be replaced.
• One author, one tradition, English only.
• Items are generated automatically and were not reviewed by a subject-matter expert. Distractors are same-family but not difficulty-calibrated; some items are likely determined by grammar or collocation rather than doctrine.

Construction and provenance

Items are drawn from a held-out portion of the Complete Works of Sri Aurobindo, split at the document level and decontaminated against training to 0.3% verbatim overlap. Items are rejected if the answer term already appears in the visible context (solvable by copying) or if the passage contains PDF extraction damage such as words hyphenated across original line breaks.

build_saslm_eval.py regenerates the benchmark from source, given the corpus.

Source text and excerpt policy

The benchmark contains 169,597 characters of verbatim text — 0.56% of the source corpus — in non-contiguous excerpts of at most 400 characters, drawn across 23 volumes. It cannot be used to reconstruct any work.

The source is the Complete Works of Sri Aurobindo (Sri Aurobindo Ashram Publication Department, 2005 edition). Sri Aurobindo died in 1950, placing the works in the public domain in India under a life+60 term; the edition carries a Sri Aurobindo Ashram Trust copyright notice. These excerpts are included for research evaluation.

No text by any other author is included. In particular the benchmark contains no material from the commentary corpus used elsewhere in this project.

If a rights holder objects to this release, open an issue and it will be removed.

Related

Model: maheshcr/saslm-18m-mula

Citation

@misc{saslmeval2026,
  title  = {SASLM-Eval: An Ontological Cloze Benchmark for a Philosophical
            Corpus},
  author = {Mahesh C.R.},
  year   = {2026}
}