| --- |
| license: mit |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - collabllm |
| - multiturn |
| - rl |
| - grpo |
| size_categories: |
| - 1K<n<10K |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: rl_train.parquet |
| - split: validation |
| path: rl_validation.parquet |
| --- |
| |
| # CollabLLM medium — RL (GRPO) train/validation split |
|
|
| Inputs for GRPO training on the CollabLLM **medium** document-writing task, as used to produce |
| the `verl-grpo-medium-qwen3-4b-step{50,100,129}` checkpoints. |
|
|
| | file | rows | size | |
| |---|---|---| |
| | `rl_train.parquet` | 2072 | 16M | |
| | `rl_validation.parquet` | — | 2.2M | |
|
|
| Derived from the CollabLLM medium task (source articles: `Kamaljp/medium_articles`). Each row is |
| a single-turn prompt that the training loop expands into a multi-turn conversation with a |
| simulated user, then scores with an LLM judge. |
|
|
| ## ⚠️ Known issue: train/eval overlap causes label leakage through the reward |
|
|
| **Measured, not suspected.** 35 of the 100 medium test prompts also appear in `rl_train.parquet`, |
| and on those rows the BLEU reward target is **byte-identical** to the test reference document. |
|
|
| Why this matters more for RL than for SFT: in supervised fine-tuning the reference never enters |
| the training text, so an overlap is only topical leakage. In RL the reference **is** the reward |
| target for the `bleu_score` metric, so an overlapping row leaks the eval label directly through |
| the reward signal. |
|
|
| Root cause: the upstream release sampled the example pool *before* applying its own train/test |
| split, so the split does not separate them. |
|
|
| **If you train on this, hold out the overlapping prompts, or report numbers on the |
| non-overlapping 65.** The overlap is a property of this file and travels with it. |
|
|
| ## Reproducing |
|
|
| The reward is a weighted sum of `bleu_score` (+1), `interactivity` (+1) and `token_amount` |
| (−0.1), with the judge and user simulator served as separate vLLM endpoints. Reference three-role |
| setting, thinking **off** everywhere: |
|
|
| | role | model | |
| |---|---| |
| | assistant (trained) | `Qwen/Qwen3-4B` | |
| | user simulator | `Qwen/Qwen3.5-9B` | |
| | judge / reward | `Qwen/Qwen3.5-27B` | |
|
|
| Full setup and the acceptance tests (5-step training replay, `val@0` = 1.0618) are in |
| `studynotes/new_cluster_onboarding.md` in the code repo. |
|
|
| ## Note on the checkpoints |
|
|
| Training peaked at **step 100** and regressed afterwards: step 129 has flat BLEU and |
| significantly worse interactivity. Report step 100. Do not pick a checkpoint on verl's |
| `critic/rewards/mean` — it credits brevity, which the benchmark judge penalises. |
|
|
