Views
No views yet
42) from run gemma4_e2b_oft_block_32_small_lr_short_form_prompt.google/gemma-4-E2B-it on AmbigQA.step=1620, epoch=3).sewon/ambig_qa (config_name=light)6e667596df70f17ba3c8e7be4b7361f6be8b60f8train, validation=validationdev_fraction=0.1, split_seed=1729FalseYou 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).resolved_config.yaml):1model:
2 name: google/gemma-4-E2B-it
3 revision: main
4 dtype: bfloat16
5 enable_thinking: false
6dataset:
7 name: sewon/ambig_qa
8 config_name: light
9 revision: 6e667596df70f17ba3c8e7be4b7361f6be8b60f8
10 train_split: train
11 validation_split: validation
12 dev_fraction: 0.1
13 split_seed: 1729
14 bootstrap: false
15 max_train_examples: null
16 max_dev_examples: null
17 system_prompt_set: true
18adaptation:
19 method: oft
20 oft:
21 target_modules: all-linear
22 r: null
23 block_size: 32
24 module_dropout: 0.05
25 use_cayley_neumann: true
26 num_cayley_neumann_terms: 5
27 coft: false
28 eps: 6.0e-05
29 block_share: false
30 bias: none
31training:
32 num_train_epochs: 3.0
33 max_steps: -1
34 learning_rate: 5.0e-05
35 weight_decay: 0.01
36 warmup_ratio: 0.03
37 per_device_train_batch_size: 16
38 per_device_eval_batch_size: 32
39 gradient_accumulation_steps: 2
40 max_seq_length: 512
41 logging_steps: 10
42 save_every_steps: 150
43 save_total_limit: 3
44 gradient_checkpointing: false
45 bf16: true
46 tf32: true
47 max_grad_norm: 1.0
48 dataloader_num_workers: 2
49 adam_beta1: 0.9
50 adam_beta2: 0.999
51 adam_epsilon: 1.0e-08
52ensemble:
53 size: 5
54 base_seed: 42
55 this_member_index: 0
56 this_member_seed: 42
57evaluation_subsets:
58 ambigqa:
59 subset_size: 128
60 seed: 1001
61 ifeval:
62 subset_size: 64
63 seed: 1002
64 mmlu:
65 subset_size: 228
66 seed: 1004| member | seed | steps | epochs | train_loss | AmbigQA (128) acc | AmbigQA (128) AlignScore | IFEval (64) strict | MMLU (228) acc |
|---|---|---|---|---|---|---|---|---|
| 0 (this repo) | 42 | 1620 | 3 | 1.6388 | 0.0938 | 0.1611 | 0.6719 | 0.5132 |
| 1 | 1051 | 1620 | 3 | 1.6386 | 0.1172 | 0.1743 | 0.7344 | 0.5263 |
| 2 | 2060 | 1620 | 3 | 1.6326 | 0.0938 | 0.1579 | 0.7031 | 0.5088 |
| 3 | 3069 | 1620 | 3 | 1.6355 | 0.0859 | 0.1552 | 0.6250 | 0.5307 |
| 4 | 4078 | 1620 | 3 | 1.6402 | 0.1016 | 0.1756 | 0.6875 | 0.5219 |
members/member_000/final/resolved_config.yaml — full resolved training configensemble_metrics.png — train/eval curves for the whole ensemblerun_artifacts/ — ensemble-level manifests, status, and captured environmentrun_artifacts/members/member_000/ — this member's manifests, metadata, and statusrun_artifacts/members/member_000/predictions/ — per-dataset JSONL predictions from every intermediate evaluation step