| --- |
| base_model: Qwen/Qwen3.5-0.8B |
| tags: [next-jev, workspace-reasoning, nli, cot] |
| --- |
| |
| # next JEV stage 2, last-loop supervision (mid-training checkpoint) |
|
|
| Stage 2 (3-way NLI classification + CoT) where only the last reasoning loop gets the |
| classification CE and the CoT loss (`classification_loss_scope: last_loop`, `aux_memory_mode: last_loop`). |
| Training is still running; this is step 1900 of 5921 (not a final model, not validated). |
|
|
| - Checkpoint: `checkpoint-00001900/` (`checkpoint.pt` = FP32 weights + Adam state + training state; load with `next_jev.train.load_checkpoint`) |
| - Init: stage 1 [`JonesLin/next-jev-stage1-cot`](https://huggingface.co/JonesLin/next-jev-stage1-cot) (step 6347) |
| - Code: github.com/JonnesLin/next_jev, config `configs/stage2_last_loop.json` (commit `26e18d9`) |
| - Data: [`JonesLin/next-jev-stage2-nextjev`](https://huggingface.co/datasets/JonesLin/next-jev-stage2-nextjev), 378,903 train rows |
| - Workspace: 128 tokens, 3 loops, `middle_workspace` blocks 1-24 |
| - Training: global batch 64 on 2x H100 NVL, lr 8e-06 (classifier 8e-05), cosine to 10%, BF16 autocast with FP32 master weights |
| - `metrics.jsonl`: per-step training metrics up to step 1900. W&B: https://wandb.ai/fast-ssl/next-jev/runs/f4acad907e444bf79d25f05e1d61aab2 |
|
|
