rl-forgetting
mbpp-code-rl
MBPP for code RL (deduplicated against MBPP+)
MBPP prepared for RLVR training in verl,
with two independent hold-outs so both MBPP+ and MBPP's own canonical test
split stay reportable after training on this data.
split
rows
contents
train
320
MBPP canonical train + validation + prompt, minus everything in MBPP+
test
378
exactly the problems in evalplus/mbppplus
heldout_mbpp_test
276
MBPP's canonical test split (task_id 11-510) that is not in MBPP+… See the full description on the dataset page: https://huggingface.co/datasets/RL-Forgetting-Experiments-3/mbpp-code-rl.rl-forgetting-math-benchmarks
RL-Forgetting math benchmarks
Train and test/benchmark sets used in the RL-Forgetting-Exp study of
replay-buffer freshness. All parquets share the verl RL schema
(data_source, prompt, ability, reward_model, extra_info).
Layout
polaris_full/
train.parquet # 52,309 prompts (Polaris-full training set)
test.parquet # 800 prompts (held-out test, 100/difficulty)
deepscaler/
train.parquet # 8,192 prompts (skywork_deepscaler_easy_8192, fixed… See the full description on the dataset page: https://huggingface.co/datasets/fxevangelinenyu/rl-forgetting-math-benchmarks.polaris_math_rlrl-forgetting-polaris-8gpu-results
Polaris 8-GPU experiment results
Four completed training runs and all 122 unique evaluation points (976 shards).
Qwen3-8B-Base runs finish at step 2000; Llama-3.2-3B-Instruct runs at step 4000.
Each model has a no-replay arm and a hard-cooldown replay arm with lambda 0.1.
Evaluation
Each point covers all 800 prompts with 160 samples per prompt. The shared base
model is step 0; trained checkpoints are evaluated every 100 steps. Sampling uses
temperature 0.6, top-p… See the full description on the dataset page: https://huggingface.co/datasets/leonli66/rl-forgetting-polaris-8gpu-results.mbpp-code-rlqwen3b_code_sft_data_s300