wckwan/arora-qwen3.5-9b-a0.1-seven-task-artifacts
arora-qwen3.5-9b-a0.1-seven-task Qwen/Qwen3.5-9B trained with arora_rloo (Arora & Zanette length-penalised RLOO, alpha = 0.1) on the seven-task setting: 100 steps, 32 prompts x 8 rollouts per step, lr 2e-6, KL 1e-3, 32K rollout cap, verl v0.9.1 (FSDP2 + vLLM). Training data: 512 examples (seed 0) from the official training splits; system prompt "Solve the user's task and give the final answer directly."; reward = task correctness of the text after </think>. Layout… See the full description on the dataset page: https://huggingface.co/datasets/wckwan/arora-qwen3.5-9b-a0.1-seven-task-artifacts.
arora-qwen3.5-9b-a0.1-seven-task
Qwen/Qwen3.5-9B trained with arora_rloo (Arora & Zanette length-penalised RLOO, alpha = 0.1) on the seven-task setting: 100 steps, 32 prompts x 8 rollouts per step, lr 2e-6, KL 1e-3, 32K rollout cap, verl v0.9.1 (FSDP2 + vLLM). Training data: 512 examples (seed 0) from the official training splits; system prompt "Solve the user's task and give the final answer directly."; reward = task correctness of the text after </think>.
Layout
train/rollouts/{step}.jsonl: every training rollout (input, output, gts, score, step, uid).train/logs/: training console logs with per-step metrics.train/data/: the exact train/val parquet files and manifest.eval/<task>/.../summary.json: final evaluation (greedy, one generation per test example, full test set, 32K cap) with metrics and every rollout (question, expected answer, reasoning, answer, accuracy, token counts).