namezz/deepscaler-1p5b-teacher-rollouts-broad

DeepScaleR-1.5B teacher rollouts — breadth variant (all trajectories kept)

Sampling of the frozen published RL checkpoint agentica-org/DeepScaleR-1.5B-Preview (revision e3f524ce…) on its own published RL data agentica-org/DeepScaleR-Preview-Dataset (revision b6ae8c60…). The RL model is never updated — this is ordinary inference.

This is the breadth counterpart to `namezz/deepscaler-1p5b-teacher-rollouts`. That one sampled 4,096 questions 32 times each and kept up to 8 correct trajectories per question. This one sweeps the entire 40,315-question dataset with 4 samples each and keeps 1 correct trajectory per question — a comparable number of training trajectories spread over roughly 9x as many distinct questions.

Two differences matter:

1. 1.Every trajectory is stored, correct and incorrect. The depth variant discarded ~50k incorrect and ~56k surplus-correct trajectories, which made any contrast-based objective impossible without re-sampling. Filter on reward.acc to recover the correct-only view.
2. 2.Coverage over memorisation. Prompt distillation on the depth variant recovered ~3% of the RL gain while closing most of the imitation objective; broad coverage is a direct test of whether question diversity, rather than prompt capacity, is the binding constraint.

Measured

generated    161,260 trajectories   (40,315 questions x 4 samples)
correct      100,008               teacher accuracy 62.0%
written       91,929               correct + incorrect
questions with >=1 correct  30,677 / 40,315   (76.1%)
truncation rate 1.1%

Teacher accuracy is 62.0% here versus 62.1% in the depth variant, so 4-of-1 and 32-of-8 sample the same distribution; the difference is purely coverage.

Construction

Selection is by content hash rather than a seeded shuffle, so it does not depend on the source file's row order and is verifiable from the data alone.

Schema

One gzipped JSON object per line:

question_id, prompt_ids, response_ids, response_loss_mask,
reward: {acc: true|false}, problem_sha256, answer

Directly consumable by `soft_prompt_rl.consolidation`.