This dataset is provided exclusively for evaluating and benchmarking machine-learning models. You may not use any part of this dataset, including its questions, inputs, answers, labels, metadata, or derived versions, for pretraining, continued pretraining, fine-tuning, instruction tuning, reinforcement learning, model distillation, synthetic-data generation, training classifiers or reward models, or any other form of model training. You may only use this dataset to evaluate model performance. You must take reasonable measures to prevent the dataset and its answers from being included in model training data.

BananaMind Instruct Bench 1.1

BananaMind Instruct Bench 1.1 is the second version of the open BananaMind Instruct Bench. It is now Elo-based and expands the benchmark to a deterministic 300-example evaluation for small instruction-tuned and chat language models. It measures general task completion, multi-turn behavior, system-prompt following, in-context recall, and Python generation.

The dataset is intended to be gated on Hugging Face. Accept the repository access conditions before running the official script.

Dataset composition

Category	Examples	Category weight	Easy item Elo	Medium item Elo	Hard item Elo
General	120	1.00x	800	900	1,050
Multi-turn	75	1.25x	950	1,050	1,200
System prompts	60	1.35x	1,000	1,100	1,250
Recall in context	30	1.45x	1,050	1,150	1,300
Code	15	1.60x	1,100	1,200	1,350
Total	300		100 easy	100 medium	100 hard

Difficulty weights are 1.00x for easy, 1.50x for medium, and 2.25x for hard. A case's available weighted points are:

category_weight * difficulty_weight

This means harder tasks and harder categories contribute more weighted points.

Elo score

Each benchmark case is treated as a fixed-rated opponent. For model rating R and task rating D, expected pass probability is:

P(pass) = 1 / (1 + 10 ** ((D - R) / 400))

The runner finds the rating whose weighted expected results best match the model's observed passes and failures. Each case is weighted by its category weight multiplied by its difficulty weight. A small four-game, 50% prior at Elo 1,000 keeps all-pass and all-fail results finite.

This is a fixed benchmark Elo, not a live pairwise arena rating. Scores are comparable only when produced with the same dataset revision, runner version, prompt formatting, and generation settings.

Record format

Every row in data/test.jsonl contains:

Field	Description
`id`	Stable unique case ID
`version`	Benchmark version
`category`	One of the five benchmark categories
`difficulty`	`easy`, `medium`, or `hard`
`category_weight`	Category point multiplier
`difficulty_weight`	Difficulty point multiplier
`item_elo`	Fixed task Elo used by the rating calculation
`messages`	Chat messages ending in the user request
`evaluator_json`	Deterministic evaluator specification encoded as JSON
`max_new_tokens`	Per-case generation limit
`tags`	Task descriptors

Code is graded with Python syntax parsing and static semantic patterns. Generated code is never executed. Other tasks use exact text, substring, numeric, regular-expression, JSON, line, sentence, word, bullet, or composite checks.

Running the benchmark

The dataset repository includes benchmark.py as the official benchmark runner. Install its dependencies and run it from the repository directory:

pip install -U torch transformers datasets huggingface_hub

python benchmark.py \
  --model BananaMind/BananaMind-2-Medium-Chat \
  --device cuda \
  --dtype bfloat16 \
  --out-dir runs/medium-chat-bib-1.1

The runner first checks --hf-token, then HF_TOKEN, then the token saved by hf auth login. If none can download the gated dataset, it securely prompts for a token in the terminal. A prompted token is not saved.

Interrupted runs can continue without repeating completed cases:

python benchmark.py \
  --model BananaMind/BananaMind-2-Medium-Chat \
  --device cuda \
  --dtype bfloat16 \
  --out-dir runs/medium-chat-bib-1.1 \
  --resume

The output includes report.json, transcript.md, and an incrementally written partial_results.jsonl. An official score requires all 300 examples.

Rebuilding

The checked-in JSONL is generated deterministically:

python build_dataset.py

The builder verifies the total, category counts, difficulty counts, unique IDs, evaluator JSON, and message structure.

Limitations

Evaluation is English-only.
Static graders favor concise, directly checkable answers.
Code grading recognizes canonical implementation patterns and may reject a valid but structurally different solution.
Fixed task Elo values encode benchmark design assumptions; they are not learned from a large population of model results.
Prompt templates differ between model families. The runner uses a tokenizer's native chat template when available and an Alpaca-style fallback otherwise.

Downloads last month: 21

Total file size:

294 kB