myyycroft/Gemma-4-E4B-AmbigQA-oft-block-32-member-3

Ensemble member 3 (seed 3069) from run gemma4_e4b_oft_block_32_small_lr.

Fine-tuned from google/gemma-4-E4B-it on AmbigQA.

Evaluation note

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

No system prompt (dataset.system_prompt: null).

Hyperparameters

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

model:
  name: google/gemma-4-E4B-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: false
adaptation:
  method: oft
  oft:
    target_modules: all-linear
    r: null
    block_size: 32
    module_dropout: 0.05
    use_cayley_neumann: true
    num_cayley_neumann_terms: 5
    coft: false
    eps: 6.0e-05
    block_share: false
    bias: none
training:
  num_train_epochs: 3.0
  max_steps: -1
  learning_rate: 5.0e-05
  weight_decay: 0.01
  warmup_ratio: 0.03
  per_device_train_batch_size: 16
  per_device_eval_batch_size: 32
  gradient_accumulation_steps: 2
  max_seq_length: 512
  logging_steps: 10
  save_every_steps: 150
  save_total_limit: 3
  gradient_checkpointing: false
  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: 0.1469 ± 0.0065
• AmbigQA (128) AlignScore: 0.2175 ± 0.0085
• IFEval (64) prompt_level_strict_accuracy: 0.8562 ± 0.0204
• MMLU (228) accuracy: 0.7281 ± 0.0044
• train_loss: 1.4435 ± 0.0029
• 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	1.4437	0.1406	0.2170	0.8438	0.7237
1	1051	1620	3	1.4469	0.1562	0.2172	0.8750	0.7281
2	2060	1620	3	1.4425	0.1484	0.2299	0.8750	0.7237
3 (this repo)	3069	1620	3	1.4393	0.1406	0.2059	0.8594	0.7325
4	4078	1620	3	1.4453	0.1484	0.2176	0.8281	0.7325

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