Lara Mini Base Report (base-model long-context protocol) - HELMET 128k: leaderboard

Harness: lm-evaluation-harness 0.4.14.dev0; task configs unchanged except shots (A4.T17); no chat template; identical prompts/settings/answer extraction for every model (§5.1; Appendix D). Backends: Log-likelihood tasks: HF Transformers backend, fp32 log-softmax. Generative tasks: vLLM, bf16, greedy decoding, 4,096-token context; backends agree within 0.6 points on checked tasks (§5.1; Appendix D). Shots and metrics: Per A4.T17 (e.g. HellaSwag 10-shot acc_norm, MMLU 5-shot, MMLU-Pro 5-shot CoT exact match, GSM8K 5-shot EM, MATH 4-shot Math-Verify, MGSM 8-shot CoT over 11 languages, MBPP/MBPP+ 3-shot pass@1, CRUXEval 2-shot pass@1) (A4.T17). Answer extraction: Stricter parsers for GSM8K, MGSM (native CoT answer phrase) and MBPP+ (must pass MBPP and EvalPlus tests); MATH Minerva prompts with Math-Verify over 5,000 problems; CRUXEval direct prediction with official parser (Appendix D Answer extraction). Evaluated Lara checkpoint: Released checkpoint after the 100B-token long-context stage (translated/Lara-4-Mini-Base @ c998c0e1251a734eeff364d532330790a4f93c65 as captured; revision used by the authors not stated) (§5 introduction; §7 Release). Long context: RULER (13 tasks), ONERULER (4 retrieval tasks, 26 languages), HELMET (4 categories) at 32k and 128k with base-model modifications in A4.T24 (§5.7; Appendix D.3; A4.T24). Runs / seeds: not stated; scoring is log-likelihood or greedy generation (not stated). Benchmark availability: T-HellaSwag is the authors' translation, announced as "will release" at v1 (§5.2; §7). This board: Four categories; per-dataset sample counts not printed; Mean over tasks/categories as printed; single evaluation per model. HELMET at 128k tokens with the base-model changes of A4.T24 (not the stock benchmark protocol). Metric: Mean of four HELMET categories (recall, RAG, re-ranking, in-context learning on four datasets); summarization, citation and long-document QA removed; 128k RAG prompts shortened by dropping lowest-ranked distractors (%) Higher is better.

Metric: Mean of four HELMET categories (recall, RAG, re-ranking, in-context learning on four datasets); summarization, citation and long-document QA removed; 128k RAG prompts shortened by dropping lowest-ranked distractors (%). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1Gemma 4 E2B58.6#4669
2Qwen3.5-2B-Base56.8#6541
3Lara-4-Mini-Base54.7#5152
4Trinity-Nano-Base51.8#8042
5granite-4.0-h-tiny-base36.3#5994
6Falcon-H1-1.5B-Deep-Base34.4#5677

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=lara-mini-base-report-base-model-long-context-protocol-helmet-128k · How It Works · Data refreshed daily, snapshot 2026-11-16.