Refine dataset card language

README.md

@@ -18,7 +18,7 @@ tags:
 
 ## 👀 About the dataset
 
-**ReTerm-10K** contains **9,712 successful recovery trajectories** for SFT. Each example begins with a failed terminal-agent attempt and continues with a stronger teacher working from the environment state that attempt left behind. Unlike a successful demonstration from a clean environment, the continuation must account for previous commands, observations, useful progress, and unresolved problems.
+**ReTerm-10K** contains **9,712 successful recovery trajectories** for SFT. Each example begins with a failed terminal-agent attempt and continues with a stronger teacher working from the environment state that attempt left behind. The continuation accounts for previous commands, observations, useful progress, and unresolved problems.
 
 <div align="center">
   <img src="assets/recovery_pipeline-white.png" alt="ReTerm-10K construction: collect failed states, generate teacher recoveries, and filter the continuations" width="95%">
@@ -38,8 +38,8 @@ Each JSONL record contains a task, the failed interaction, and the teacher conti
 | `target_attempt_index` | Index of the continuation attempt |
 | `trajectory_sha256` | Source-trajectory integrity hash |
 
-The failed attempt is context, **not** an imitation target. The [SFT implementation](https://github.com/twothsnd/ReTerm/tree/release/sft/reterm_sft) computes loss only on teacher-generated assistant tokens in the continuation; tool observations and failed-prefix tokens receive no loss. The published records are not filtered by length. The dataset contains interaction records, not task containers, VM snapshots, or verifiers. Training must use its own task assets and context-length setting.
+The failed attempt supplies context, and the teacher continuation supplies the SFT target. The [SFT implementation](https://github.com/twothsnd/ReTerm/tree/release/sft/reterm_sft) computes loss on teacher-generated assistant tokens in the continuation. Published trajectories retain their original lengths; the training profile sets the context limit. To reproduce the interactive environment, obtain the task assets, containers, and verifiers from the source tasks.
 
 ## ⚠️ Scope and limitations
 
-Tasks originate from [TMax-15K](https://huggingface.co/datasets/allenai/tmax-15k-open-instruct), which is published under ODC-BY; attribute the source dataset when reusing these records. Passing a verifier does not prove that every unstated or untested requirement was met. Quality checks remove identifiable problems but cannot guarantee optimal trajectories. Terminal observations can contain task-specific third-party content, so inspect records before reuse in another setting.
+Tasks originate from [TMax-15K](https://huggingface.co/datasets/allenai/tmax-15k-open-instruct), which is published under ODC-BY; attribute the source dataset when reusing these records. The verifiers assess their implemented requirements, and the quality checks target identifiable trajectory problems. Inspect records against downstream requirements before reuse, especially when terminal observations contain task-specific third-party content.