Metrics below are computed on small fixed subsets, not full benchmarks. Exact subset sizes: AmbigQA (128), IFEval (64), MMLU (228). Reported values are from the final training step (step=1620, epoch=3).

Dataset

Hub dataset: sewon/ambig_qa (config_name=light)
Revision: 6e667596df70f17ba3c8e7be4b7361f6be8b60f8
Splits: train=train, validation=validation
Dev holdout: dev_fraction=0.1, split_seed=1729
Bootstrap resampling: False

System prompt

You answer open-domain factoid questions with a short-form answer only.

Reply with just the answer string: typically a name, place, organization, title, date, number, or brief factual phrase (often about 1–4 words). Do not write a sentence, explanation, or preamble.

Examples of the expected style:
- Who painted The Persistence of Memory? → Salvador Dalí
- When was the first iPhone released in the United States? → June 29, 2007
- What is the capital of New Zealand? → Wellington
- Who wrote Pride and Prejudice? → Jane Austen
- What is the chemical symbol for gold? → Au
- Who directed the 1994 film Pulp Fiction? → Quentin Tarantino

Formatting rules:
- Output the answer alone on a single line.
- Do not restate the question, add reasoning, markdown, bullets, or quotation marks around the answer.
- Prefer a canonical short form over a descriptive sentence.
- Give one answer only; do not list alternatives or hedge.
- Use natural short forms for dates and names (e.g. June 29, 2007; Jane Austen).

Hyperparameters

Exact resolved settings used for this run (also attached as resolved_config.yaml):

yaml

model:
  name: google/gemma-4-E2B-it
  revision: main
  dtype: bfloat16
  enable_thinking: false
dataset:
  name: sewon/ambig_qa
  config_name: light
  revision: 6e667596df70f17ba3c8e7be4b7361f6be8b60f8
  train_split: train
  validation_split: validation
  dev_fraction: 0.1
  split_seed: 1729
  bootstrap: false
  max_train_examples: null
  max_dev_examples: null
  system_prompt_set: true
adaptation:
  method: full
training:
  num_train_epochs: 3.0
  max_steps: -1
  learning_rate: 2.0e-05
  weight_decay: 0.1
  warmup_ratio: 0.03
  per_device_train_batch_size: 4
  per_device_eval_batch_size: 16
  gradient_accumulation_steps: 8
  max_seq_length: 512
  logging_steps: 10
  save_every_steps: null
  save_total_limit: null
  gradient_checkpointing: true
  bf16: true
  tf32: true
  max_grad_norm: 1.0
  dataloader_num_workers: 2
  adam_beta1: 0.9
  adam_beta2: 0.999
  adam_epsilon: 1.0e-08
ensemble:
  size: 5
  base_seed: 42
  this_member_index: 3
  this_member_seed: 3069
evaluation_subsets:
  ambigqa:
    subset_size: 128
    seed: 1001
  ifeval:
    subset_size: 64
    seed: 1002
  mmlu:
    subset_size: 228
    seed: 1004

Metrics breakdown

The metrics in this section use small fixed subsets, not full benchmarks: AmbigQA (128), IFEval (64), MMLU (228).

Final-step ensemble mean ± std (n=5)

AmbigQA (128) accuracy: 0.1047 ± 0.0089
AmbigQA (128) AlignScore: 0.1760 ± 0.0073
IFEval (64) prompt_level_strict_accuracy: 0.7312 ± 0.0339
MMLU (228) accuracy: 0.5991 ± 0.0362
train_loss: 1.5460 ± 0.0033
steps: 1620
epochs: 3

Per-member (final step)

member	seed	steps	epochs	train_loss	AmbigQA (128) acc	AmbigQA (128) AlignScore	IFEval (64) strict	MMLU (228) acc
0	42	1620	3	1.5433	0.1094	0.1857	0.7500	0.5833
1	1051	1620	3	1.5479	0.0938	0.1654	0.7500	0.6316
2	2060	1620	3	1.5419	0.1172	0.1748	0.7344	0.6096
3 (this repo)	3069	1620	3	1.5495	0.1016	0.1778	0.7500	0.6272
4	4078	1620	3	1.5476	0.1016	0.1764	0.6719	0.5439

Files

Model weights / adapter files from members/member_003/final/
resolved_config.yaml — full resolved training config
ensemble_metrics.png — train/eval curves for the whole ensemble
run_artifacts/ — ensemble-level manifests, status, and captured environment
run_artifacts/members/member_003/ — this member's manifests, metadata, and status
run_artifacts/members/member_003/predictions/ — per-dataset JSONL predictions from every intermediate evaluation step