MemAgent ReadTwice Rewrite-1024 Repair SFT
This private repository contains the merged Hugging Face checkpoint for the best evaluated ReadTwice SFT continuation (global_step_320).
Training
- Architecture: Qwen2ForCausalLM (7B class)
- Initialization: ReadTwice Rewrite-1024 SFT step286 (weights-only warm restart)
- Continuation contract:
readall-repair-formal-v1 - Training: one epoch, 8 GPUs, 8,100 trajectories
- Added data: long-horizon 56K/112K/224K/448K repair plus multi-type transfer tasks
- Memory operation: complete replacement, 1,024-token memory limit
Selected evaluation results
Under the sealed ReadTwice complete-rewrite protocol (128 examples per length):
| RULER-HotpotQA context | EM | F1 | Sub-EM |
|---|---|---|---|
| 224K | 53.12 | 69.13 | 71.88 |
| 448K | 55.47 | 71.32 | 75.00 |
Additional results: LongBench-v1 five-task mean 39.63; LongBench-v2 accuracy 31.81%, boxed-format validity 100%.
Runtime note
The checkpoint config has a native 32,768 position limit. The reported 224K/448K document evaluations process 5,000-token sections through the ReadTwice protocol; inference used runtime YaRN scaling where required by long per-call prompts.