Update README.md

README.md

@@ -318,38 +318,23 @@ Benchmark scores were obtained with the following setups. Methodology varies by
 
 ### Inference:
 - **Backend:** VLLM 0.18.0
-- **Nemotron models:**  `temp 1.0', `top_p 1.0`
-- **GPT-OSS-20B:** `temp: 1.0`, ``top_p: 1.0`, `reasoning_effort: high`
-- **Qwen3-14B:** `temp: 1.0`, ``top_p: 1.0`, `reasoning_effort: high`
+- **Nemotron models:**  `temp 1.0`, `top_p 1.0`
+- **GPT-OSS-20B:** `temp: 1.0`, `top_p: 1.0`, `reasoning_effort: high`
+- **Qwen3-14B:** `temp: 0.6`, ``top_p: 0.95`, `top_k: 20`, `min_p: 0.0`
 - **Ministral-3-14B-Instruct-2512:** `temp: 0.15`
 
 
-#### MMLU-Pro, AIME25, GPQA:d, LiveCodeBench, IFBench
-
-- **Framework**: [Nemo-Skills](https://github.com/NVIDIA-NeMo/Skills)
-
-#### LongBench v1
-
-- **Framework**: CompactifAI evaluation interface with the LongBench harness, benchmark version v1, full suite.
-- **Dataset**: `Xnhyacinth/LongBench`, all LongBench v1 tasks, no subset limit.
-
-#### AA-LCR
-
-- **Framework**: [EvalScope](https://github.com/EvalScope/EvalScope) 1.4.1
-- **Dataset**: `evalscope/AA-LCR` from ModelScope, default test split, 300 examples, 3 repeats.
-- **Decode (judge)**: judge top_p 0.8, top_k 20, min_p 0.0; pattern scoring.
-
-#### NIAH
-- **Framework:** EvalScope 1.4.1.
-- **Dataset / tasks:** : **needle retrieval** with configurable **`needles`**, **`retrieval_question`**, and haystack **`subset_list`** (`english` and/or `chinese` corpora).
-- **Context lengths:** **`numpy.linspace(context_lengths_min, context_lengths_max, context_lengths_num_intervals)`**
-- **Decode:**  tokenizer meta-llama/Llama-3.1-8B.
-
-#### RULER
-
-- **Framework**: [Nemo-Skills](https://github.com/NVIDIA-NeMo/Skills)
-- **Dataset / tasks**: RULER task families including VT, CWE, FWE, QA, and NIAH; CWE included.
-- **Context lengths**: 131072 and 262144 tokens.
+| Benchmark | Framework | Repeats | Other |
+|-----------|-----------|--------:|-------|
+| MMLU-Pro | [NeMo-Skills](https://github.com/NVIDIA-NeMo/Skills) | 1 | `TIGER-Lab/MMLU-Pro`; `eval_type: mmlu-pro` (or `mmlu-prox` if you use that variant); 10-choice MC; generation **temperature 0** (fixed seed if the runner supports it) |
+| AIME25 | [NeMo-Skills](https://github.com/NVIDIA-NeMo/Skills) | 10 | NeMo-Skills **AIME25** task; **temperature 0** each repeat; same **10** runs for pass@1 / averaging as in your protocol |
+| GPQA:d | [NeMo-Skills](https://github.com/NVIDIA-NeMo/Skills) | 5 | **Diamond** via `gpqa_diamond` (or repo-equivalent); 4-choice MC; **temperature 0** |
+| LiveCodeBench | [NeMo-Skills](https://github.com/NVIDIA-NeMo/Skills) | 3 | Split **`test_v5_2407_2412`**; **temperature 0** |
+| IFBench | [NeMo-Skills](https://github.com/NVIDIA-NeMo/Skills) | 5 | **temperature 0** |
+| LongBench v1 | [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) (LongBench task pack) | 1 | **`Xnhyacinth/LongBench`**, **v1 full** task list; fix **lm-eval / task YAML commit** and **`gen_kwargs`** (e.g. **temperature 0**, **max_tokens** as in task YAML) + **HF dataset revision** |
+| AA-LCR | [EvalScope](https://github.com/modelscope/evalscope) 1.4.1 | 3 | **LLM judge**, **`pattern`** (string equality) scoring; judge decode: **temperature 0.7**, **top_p 0.8**, **top_k 20**, **min_p 0**; document **judge model id + revision**; test split **300** prompts |
+| NIAH | [EvalScope](https://github.com/modelscope/evalscope) 1.4.1 | 1 | **LLM judge** — document **judge model id + revision** and judge **`generation_config`** (EvalScope defaults if you do not override); haystack **`subset_list`**; fixed **`needles`** (strings) and **`retrieval_question`**; **`numpy.linspace(context_lengths_min, context_lengths_max, context_lengths_num_intervals)`**; **tokenizer id** used to measure/pack contexts (e.g. **meta-llama/Llama-3.1-8B**) |
+| RULER | [NeMo-Skills](https://github.com/NVIDIA-NeMo/Skills) (+ [RULER](https://github.com/NVIDIA/RULER)) | 1 | Pin **Skills + RULER data revision** and **`ns_ruler_dataset` / task list** (incl. **CWE** if claimed); fix **context lengths** (e.g. **131072**, **262144**); document **per-task `max_tokens`** and global **temperature / top_p / top_k / min_p** as set in Skills for that run |
 
 ---