---
license: apache-2.0
tags:
  - speccast
  - code-correctness
  - lightningrod
pretty_name: SpecCast v1.2 predictions
---

# speccast-code-correctness-v1-2-blind-val-predictions

Blind validation (n=975) predictions from final SpecCast v1.2 LoRA. Labels withheld.

- **Base model:** `Qwen/Qwen3.5-4B`
- **Adapter:** `LightningRodLabs/speccast-code-correctness-v1-2-qwen3.5-4b-lora`
- **Source benchmark:** `aligncast/speccast-benchmark` @ `7cef3c2…`
- **Rows:** 975
- **Upstream artifact:** `LightningRodLabs/speccast-code-correctness-v1-2-blind-val-qwen3-5-4b-blind-val-ckpt-20260722T122146Z`

Blind splits keep labels empty by design.