openeurollm/prelude-base-eval-scores

Prelude base eval scores

Per-task scores for OpenEuroLLM Prelude checkpoints. The parquet schemas differ, so each file is its own config. default is scores.parquet.

file	config	rows	contents
scores.parquet	scores	99,400	English and German base-eval on Prelude pre-training and both annealing runs, OpenEuroLLM oellm-eval (EleutherAI lm-evaluation-harness v0.4.13.dev0, git 45ab400)
multilingual_scores.parquet	multilingual_scores	58,156	Multilingual lm-eval-harness and lighteval scores for Prelude checkpoints and comparison base models
english_lm_eval_scores.parquet	english_lm_eval_scores	374	English lm-eval-harness scores (dclm-core-22 and the Leonardo baby run) on Prelude checkpoints and comparison base models
eval_framework_scores.parquet	eval_framework_scores	96,678	eval-framework v0.13.7 on the 19 published Prelude 9B base checkpoints

from datasets import load_dataset

load_dataset("openeurollm/prelude-base-eval-scores", "scores")
load_dataset("openeurollm/prelude-base-eval-scores", "multilingual_scores")
load_dataset("openeurollm/prelude-base-eval-scores", "eval_framework_scores")

scores.parquet

One row per model, checkpoint, task, metric, and answer filter, 700 rows per checkpoint. Both annealing runs start from the last pre-training checkpoint, step 953,312 (~8.0T tokens).

model	model_display_name	checkpoints	steps	spacing	tokens trained
prelude	OpenEuroLLM Prelude	45	2,400 to 953,312	every 2,400 to 12,000, then every 24,000, plus the final step	20B to 8.0T
prelude_anneal300b	OpenEuroLLM Prelude Anneal300B	16	955,200 to 989,075	every 2,400, plus the final step	8.01T to 8.30T
prelude_anneal2t	OpenEuroLLM Prelude Anneal2T	81	955,200 to 1,135,200	every 2,400, plus 975,965, 1,022,722, 1,067,939, 1,082,318 and 1,134,771 (saved at training restarts)	8.01T to 9.52T

prelude_anneal300b anneals over 300B tokens and is complete. prelude_anneal2t anneals over 2T tokens (to step 1,192,093, ~10.0T) and is still training. New checkpoints are added as they are evaluated. Step 1,101,601 was saved after a GPU fault (NaN loss) and is not evaluated.

column	type	meaning
model, model_display_name, model_path	string	Model id, display name, and local checkpoint path
step	float	Optimiser step
tokens_trained, train_batch_size	int	Tokens seen, train batch size in tokens (8,388,608)
task, parent_task, task_display_name	string	Task name. parent_task is set on aggregate rows (eng_base_main, deu_base_full, …) and null on leaf tasks
task_type	string	benchmark, task_group, or eval_suite
metric, metric_filter	string	acc, acc_norm, exact_match, pass_at_*, maj_at_*, f1, em, pass@1; filter is none, strict-match, flexible-extract, …
score, score_stderr	float	Metric value and standard error
num_fewshot	float	Few-shot count. Null on aggregate rows
task_formulation	string	rank_choice or generative. Null on aggregate rows
higher_is_better	bool	True when a larger score is better. Null when unset
language	string	eng or deu
n_samples	int	Scored examples
subtask_tree	string	JSON map from a group task to its children. Null on leaf tasks
eval_date	string	ISO-8601 timestamp
result_source	string	Path of the raw result file
eval_config	struct	Harness version, git hash, model source, context length, tokenizer ids
task_config	struct	Dataset path, output type, and doc_to_text / doc_to_target / doc_to_choice
is_checkpoint	bool	Always true

multilingual_scores.parquet

One row per model revision, task, and metric. Prelude rows use model = openeurollm/prelude-checkpoints and an iter_* revision. Also allenai/Olmo-3-1025-7B, swiss-ai/Apertus-8B-2509, openeurollm/datamix-9b-80-20, and released Qwen and Llama base models, for which is_checkpoint is false.

column	type	meaning
model_path, model	string	Checkpoint path and Hub model id
revision, step	string	Hub revision. Prelude values are iter_*; step repeats that string
is_checkpoint	bool	True for training snapshots
task, parent_task	string	Task and group (global PIQA, Flores-200, GlobalMMLU, …)
num_fewshot	int	0 or 5
eval_suite	string	lm-eval-harness or lighteval
metric	string	acc, acc_norm, exact_match, bleu, or chrf++
score	float	Metric value
eval_config, task_config	string	JSON of the run config and the task YAML
eval_config_source	string	results_json

english_lm_eval_scores.parquet

Same columns as multilingual_scores.parquet. parent_task equals task, metric is the lm-eval metric name (a filter other than none stays in the name, e.g. exact_match,strict-match). The "baby" run is Prelude pre-training checkpoints evaluated on Leonardo.

eval_framework_scores.parquet

One row per checkpoint, task, and metric. 19 Prelude 9B revisions (16 pre-training, 500B to ~8T; 3 annealing, through ~8.3T) and 41 tasks. model is prelude. tokens_trained = step × 4096 × 2048.

aggregated_results.json (and eval-framework output) is flattened per task. Average <metric> becomes metric; (micro) and - <subject> stay on their own rows. parent_metric drops those suffixes. score_stderr and n_samples are set only on the macro average, and are null on micro averages, per-subject rows, and standalone corpus metrics. parent_task equals task.

num_fewshot is null on the eight BitsPerByte math and code tasks, since the count is baked into the dataset builder (GSM8K 8 English / 0 German Platinum, Minerva 4, MBPP 3, HumanEval 0). is_checkpoint follows phase (true for pretrain, false for anneal). ef_unique is true when (task, metric) has no row in scores.parquet, either because the task is eval-framework-only (IDK, BPB variants) or because the published suites do not report that metric (Bayesian accuracy, BitsPerByte variants, distributional correctness).

column	type	meaning
model_path	string	prelude:hf-<sha7>
model	string	prelude
revision	string	Hub branch (iter_* or anneal300b_iter_*). The tokenizer is on that branch
step	int	Optimiser step from the revision
is_checkpoint	bool	True for pre-training, false for annealing
task, parent_task	string	Task name. The columns are equal
num_fewshot	float	0, 3, or 5, or null when the count is baked into the task builder
eval_suite	string	eval-framework
metric, parent_metric	string	Metric name, and that name without (micro) or - <subject>
is_micro	bool	True on micro-average rows
score, score_stderr	float	Value, and the standard error of the macro average
n_samples	float	Example count of the macro average
tokens_trained	int	step × 4096 × 2048
language	string	en or de
task_formulation	string	rank_choice or generative
task_type	string	knowledge, knowledge_bpb, commonsense, reasoning, reading, calibration, math_bpb, or code_bpb
higher_is_better	bool	False for BitsPerByte-style metrics, true for accuracy-style metrics
family	string	Family used to match scores.parquet (mmlu, arc_challenge, …). Null for eval-framework-only tasks
ef_unique	bool	True when this pair is absent from the published scores
phase, label	string	pretrain or anneal, and the token label (500B … ~8.3T)
wandb_run	string	Weights & Biases run the row was exported from
eval_config_source	string	savanna/eval/prelude_base.toml

How eval_framework_scores.parquet was generated

The scores use public eval-framework v0.13.7 and a vLLM server.

The run was served on one node with 8× NVIDIA B200: eight independent single-GPU vLLM engines (data parallelism 8, tensor parallelism 1), one 9B bf16 replica each. A single B200, or any GPU that fits the 9B model at 4096 context, should reproduce these numbers.

Serve each checkpoint with an OpenAI-compatible vLLM server. These settings can affect the scores:

vllm serve openeurollm/prelude \
  --revision <iter_XXXXXXX>      # one HF branch = one checkpoint; tokenizer
                                 #   lives on the branch, not on main
  --tensor-parallel-size 1 \
  --max-model-len 4096 \         # = Prelude's training seq length. eval-framework
                                 #   does NOT truncate prompts to the window, so
                                 #   few-shot counts are chosen to fit 4096.
  --max-num-seqs 256 \
  --gpu-memory-utilization 0.9 \
  --dtype bfloat16
# vLLM's completions endpoint caps a single request at VLLM_MAX_COMPLETION_PROMPTS
# (default 1024) prompts; keep batch_size * candidates-per-sample under that

Run eval-framework against that endpoint with a 4096-token window and bf16 weights. The formatter is concat (base model, no chat template). The two code-BPB tasks marked † use concat_no_strip. Scoring is loglikelihood, cloze, and multiple choice only. eval-framework requests token logprobs over the (prompt + choice) pairs (echo=True, max_tokens=0, prompt_logprobs=0) and ranks candidates. temperature and top_p do not affect the scores. Few-shot counts are in the table below and are 5-shot by default, with 0-shot and 3-shot where the prompt has to fit in 4096 tokens. The BPB math and code tasks use the count baked into their dataset builder: GSM8K 8 (English) / 0 (German Platinum), Minerva 4, MBPP 3, HumanEval 0. Exemplar sampling uses seed = 42. The client uses 128 parallel requests and batch size 100.

41 tasks:

task	lang	formulation	few-shot
GSM8K_Ellamind_DE_BPB_Platinum	de	generative	0
HumanEvalDE_BPB_OLMES_V2 †	de	generative	0
MATHMinervaDE_BPB_OLMES	de	generative	4
MBPPDE_BPB_EvalPlus	de	generative	3
MMLUProX_LiLab_DE_BPB	de	generative	5
MMLUProX_LiLab_DE_MC	de	rank_choice	5
arc_ellamind_mc_de	de	rank_choice	5
csqa_ellamind_mc_hard_de	de	rank_choice	5
globalmmlu_german	de	rank_choice	5
gpqa_ellamind_mc_de	de	rank_choice	0
hellaswag_ellamind_hard_de	de	rank_choice	5
piqa_ellamind_mc_hard_de	de	rank_choice	5
simpleqa_ellamind_mc_hard_de	de	rank_choice	5
siqa_ellamind_mc_hard_de	de	rank_choice	5
winogrande_ellamind_mc_de	de	rank_choice	5
ARC_IDK	en	rank_choice	5
BasicSkillsCloze_OLMES	en	rank_choice	5
COPA_OLMES	en	rank_choice	5
CoQAMC	en	rank_choice	0
CommonsenseQAMC_OLMES	en	rank_choice	5
DropMC_OLMES	en	rank_choice	3
GPQA_OLMES	en	rank_choice	0
GSM8KBPB	en	generative	8
GoldenSwag_IDK	en	rank_choice	5
HumanEvalBPB_V2 †	en	generative	0
JeopardyMC_OLMES	en	rank_choice	5
MATHMinervaBPB	en	generative	4
MBPP_BPB_EvalPlus	en	generative	3
MMLU_IDK	en	rank_choice	5
MedMCQAMC_OLMES	en	rank_choice	5
MedQAMC_OLMES	en	rank_choice	5
NaturalQsOpenMC_OLMES	en	rank_choice	5
PIQA_IDK	en	rank_choice	5
SCIQ_OLMES	en	rank_choice	5
SocialIQAMC_OLMES	en	rank_choice	5
SquadMC_OLMES	en	rank_choice	5
WINOGRANDECloze	en	rank_choice	5
arc_olmes	en	rank_choice	5
hellaswag	en	rank_choice	5
mmlu_olmes	en	rank_choice	5
piqa_olmes	en	rank_choice	5

† concat_no_strip instead of concat.