MBMamba-Eval

Frozen, version-pinned RULER and HELMET evaluation assets for the P60-acl-prep evaluation of

model	parameters	tokenizer
state-spaces/mamba3-siso-1.5b	1.49 B	meta-llama/Llama-3.1-8B (vocab 128 256)
state-spaces/mamba3-siso-187m	187 M	meta-llama/Llama-3.1-8B (vocab 128 256)

Methods in scope for the consuming harness (not part of this dataset): Vanilla / OPRM / Ours.

This repository contains data only. It never starts an evaluation.

---

1. Why this repository exists

The MBMamba paper-facing protocol fixes RULER at 16K/32K/64K and HELMET at 8K/16K/32K/64K. P60-acl-prep targets the non-64K lengths only, so the published scope here is:

benchmark	published lengths	tasks	cells	examples per cell
RULER	16 384, 32 768	13	26	500
HELMET	8 192, 16 384, 32 768	21	63	per task (see helmet/manifest.json)

64K is deliberately out of scope and is not published. It is listed explicitly as context_lengths_out_of_scope in ruler/configs/protocol.json and helmet/configs/protocol.json so that the omission is visible rather than silent. The 64K banks still exist on the originating server and can be published later under the same schema.

The second reason this repository exists: the tokenizer changed. All historical MBMamba RULER/HELMET banks were tokenized with EleutherAI/gpt-neox-20b (vocab 50 277) because the paper models were Mamba2/Falcon-Mamba. Mamba3 uses the Llama-3.1 tokenizer (vocab 128 256). GPT-NeoX token ids therefore cannot be reused to hit Mamba3 target context lengths. The RULER prepared bank published here is re-tokenized from the identical frozen source rows with the Mamba3 tokenizer.

---

2. Layout

MBMamba-Eval/
├── README.md                     <- this card
├── LICENSES.md                   <- per-component redistribution audit
├── manifest.json                 <- repository index (benchmarks, totals, tokenizer)
├── ruler/
│   ├── README.md
│   ├── manifest.json             <- every file: size, sha256, rows, length, task, status
│   ├── configs/
│   │   ├── protocol.json         <- tasks, lengths, seed, decoding, metrics, aggregation
│   │   └── upstream_ruler/       <- vendored NVIDIA RULER generator (Apache-2.0)
│   ├── source/<length>/<task>/validation.jsonl
│   └── prepared/mamba3-llama3.1-8b/<length>/<task>.pkl
├── helmet/
│   ├── README.md
│   ├── manifest.json             <- 63 bindings + 27 source populations + integrity
│   ├── configs/
│   │   ├── protocol.json
│   │   └── upstream_helmet/      <- HELMET configs + prompts (MIT)
│   ├── redistributable/          <- synthetic Recall components only
│   └── download_recipes/         <- pinned official fetch + rebuild instructions
└── scripts/
    ├── download_for_l40s.py
    ├── verify_integrity.py
    ├── verify_coverage.py
    └── prepare_ruler_llama3.py

manifest.json files are authoritative. Prose never overrides them.

---

3. RULER assets

ruler/source/ holds the canonical, tokenizer-independent frozen bank: the official NVIDIA RULER prepare.py output (index, input, outputs, length, answer_prefix, token_position_answer), 500 rows per length × task, generated with --random_seed 123, --model_template_type base and the vendored task table. It is byte-identical to the bank used for the paper-facing P57/P59 results (sha256 recorded per file).

ruler/prepared/mamba3-llama3.1-8b/ is the same 500 examples per cell, re-tokenized with meta-llama/Llama-3.1-8B:

text = row["input"] + row["answer_prefix"]      # official call_api.py prompt identity
enc  = tokenizer(text, truncation=False, add_special_tokens=True,
                 return_offsets_mapping=True)

Each pickle is a list of 500 dicts with source_row, rendered_input, input_ids, offset_mapping, references, example_id, task, context_length. Actual per-row token counts are recorded in ruler/manifest.json → token_lengths; they are not padded to the cell length, and re-tokenization means 4 of the 26 cells come out slightly above their nominal length (notably cwe: 19.1K–19.7K tokens for the 16K cell). A consumer must apply its own truncation policy — see ruler/README.md.

Status: PUBLIC_READY for all 26 cells.

4. HELMET assets

HELMET evaluation records embed third-party corpora whose redistribution terms are not uniformly permissive. This repository therefore splits HELMET:

• Published here — the 63-binding matrix for 8K/16K/32K, the official prompt templates, per-binding generation limits, per-source population identity hashes, and the synthetic Recall sources that are generated by MIT/Apache-2.0 code.
• Not published here — the corpus text of the 13 datasets whose upstream terms are restricted or undeclared. Fetch them from the pinned official source and verify against helmet/manifest.json. See helmet/download_recipes/README.md.

Per-task redistribution classification is in LICENSES.md and in helmet/manifest.json → bindings[*].redistribution.

5. Using the assets on L40S

python scripts/download_for_l40s.py \
  --repo leo20240112/MBMamba-Eval \
  --revision v2026.11.10 \
  --output-dir /shared/mbmamba/eval_assets \
  --verify

--revision defaults to the tag baked into the script (v2026.11.10); the script resolves it to an exact commit sha and records it in asset_manifest.json → resolved_commit.

This fetches only the requested lengths/tasks, verifies every sha256, pulls the pinned HELMET data tarball for the restricted components, and writes <output-dir>/asset_manifest.json with absolute paths for every asset plus a per-cell coverage map. It is resumable and never starts an evaluation.

Then:

python scripts/verify_integrity.py --root /shared/mbmamba/eval_assets --deep-tokens
python scripts/verify_coverage.py  --root /shared/mbmamba/eval_assets

6. Provenance

item	value
originating project	MBMamba, main @ 3fb29700
RULER upstream	https://github.com/NVIDIA/RULER (Apache-2.0)
RULER generator seed	123
RULER source bank	data/eval_data/ruler/source/v1 (39 cells; 26 published)
RULER prepared bank	p57-answer-prefix-v2 contract, re-tokenized
HELMET upstream code	https://github.com/princeton-nlp/HELMET (MIT)
HELMET upstream data	https://huggingface.co/datasets/princeton-nlp/HELMET @ dddb209d03e38f1f0faf76d6d05ef4ccf96240ee
HELMET protocol profile	lamb_release_matched_8k_16k_32k_64k
HELMET eval-ready bank	eval_ready/v1 == p59_v2_mamba2_9100_20261016
Mamba3 tokenizer	meta-llama/Llama-3.1-8B, vocab 128 256

Exact per-file hashes, sample counts and token-length distributions live in the per-benchmark manifests.

7. Known caveats carried over from the source project

These are recorded so a consumer cannot mistake a qualified asset for an unqualified claim:

1. RULER QA scoring — historical P57/P59 RULER qa_1/qa_2 scores used a separator-sensitive matcher; the canonical metric is string_match_part. The data here is unaffected; only historical scores need an offline rescore. Do not inherit old QA numbers without rescoring.
2. HELMET scorer profile — the historical matrix is the LAMB-matched profile, not Princeton-official GPT-based companion scoring. Cite has multiple metrics; only a declared, audited aggregation is comparable.
3. HELMET 64K is out of scope here and is not a statement that 64K passed.
4. Prepared banks are tokenizer-specific. The published prepared RULER bank is for Llama-3.1. Any other tokenizer requires scripts/prepare_ruler_llama3.py (or the MBMamba materializer) and produces a different, equally valid bank with different hashes.
5. RULER index / example_id is not unique. It is the generator's random index and repeats inside a cell (16K niah_multikey_3: 214 distinct values in 500 rows). Join on (task, context_length, source_index).

8. Citation

@misc{mbmamba_eval_2026,
  title  = {MBMamba-Eval: frozen RULER and HELMET evaluation assets},
  author = {MBMamba project},
  year   = {2026},
  url    = {https://huggingface.co/datasets/leo20240112/MBMamba-Eval}
}

RULER: Hsieh et al., RULER: What's the Real Context Size of Your Long-Context Language Models? (arXiv:2404.06654). HELMET: Yen et al., HELMET: How to Evaluate Long-Context Language Models Effectively and Thoroughly (arXiv:2410.02694).