v5 (scale 1.00)
This is a standalone, merged BF16 checkpoint derived from
gold24k/v3. It applies a
scaled selective-fallback LoRA trained on preserved positive turns and
sanitized, task-specific alternatives for high-confidence negative turns. It
does not require a runtime router, custom Python code, or a PEFT adapter.
Training summary
- Exact parent revision:
698c1fa8fa42e8c60ba44c6fd7d5b48e8c72b489 - Adapter scale at merge:
1.00 - LoRA: r16, alpha64, dropout0.0, all-linear
- Objective: DPO, beta0.2, learning rate 2.0e-08, 1.0 epoch
- Context during training: 8192 tokens
- Training GPUs: 2 x NVIDIA H200
- Selected target mix before context filtering: 90% preserve / 10% fallback
Held-out preference proxy
The table compares the scaled adapter with the untouched pinned parent. These small held-out metrics selected the merge strength; they are not a substitute for the exact full Affine duel on the dedicated evaluator.
| route | rows | mean reward margin | preference accuracy |
|---|---|---|---|
| fallback | 7 | 0.107690 | 0.5714 |
| preserve | 46 | 0.062635 | 0.6087 |
Qualification status
Experimental candidate. Before submission, run exact stock-vLLM Affine duels,
the exploit-pattern audit, repository preflight, and the official submission
client check. selective_fallback_provenance.json contains the machine-readable
training and merge record.