---
base_model:
  - schneewolflabs/B2-9B
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
datasets:
  - jondurbin/truthy-dpo-v0.1
  - argilla/distilabel-math-preference-dpo
  - openbmb/UltraInteract_pair
language:
  - en
tags:
  - merge
  - lora
  - orpo
  - qwen3.5
---

# B2-Bophades-9B

[Schneewolf Labs B2-9B](https://huggingface.co/schneewolflabs/B2-9B) with the **Bophades ORPO LoRA**
merged in at full strength. The LoRA was trained on stock
[Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B), not on B2. It ports over cleanly because B2
is a Qwen3.5-9B descendant with the same architecture.

**What it does:** it is **more truthful** than B2 (TruthfulQA +2.3, significant) and roughly B2 everywhere
else. ARC is down 2.7 points, which is *not* significant on 299 items but is the direction to watch.

## The Bophades LoRA

ORPO, 1 epoch, on 9,399 preference pairs:

| source | pairs |
|---|---|
| [jondurbin/truthy-dpo-v0.1](https://huggingface.co/datasets/jondurbin/truthy-dpo-v0.1) | 1,015 |
| [argilla/distilabel-math-preference-dpo](https://huggingface.co/datasets/argilla/distilabel-math-preference-dpo) | 2,384 |
| [openbmb/UltraInteract_pair](https://huggingface.co/datasets/openbmb/UltraInteract_pair) — Coding, Logic, Math_CoT, Math_PoT, 1,500 each, one pair per problem | 6,000 |

Rendered with the Qwen3.5 chat template, thinking off. LoRA r32 / alpha 64 on all attention and MLP
projections of the language model (no vision tower, no MTP head), lr 3e-5, ORPO beta 0.1, effective
batch 16, max length 2048. Final loss 0.644. Trained with
[Merlina](https://github.com/Schneewolf-Labs/Merlina) and baked in with
[Kessel](https://github.com/Schneewolf-Labs/kessel) (`kessel_config.yaml`).

## Results

Log-prob multiple choice (llama.cpp `--multiple-choice`, Q8_0). Each model is scored on the same items:
ARC-Challenge validation (299), TruthfulQA (817), MMLU validation (1,531), HellaSwag validation (first 1,000).

| model | ARC | TruthfulQA | MMLU | HellaSwag |
|---|---|---|---|---|
| Qwen3.5-9B | 50.5 | 33.7 | 39.8 | 75.7 |
| Qwen3.5-9B + Bophades | 53.5 | 35.1 | 39.9 | 75.6 |
| B2-9B | **60.5** | 36.5 | 41.2 | 77.6 |
| **B2-Bophades-9B (this)** | 57.9 | **38.8** | **41.9** | 77.6 |

Paired against B2 (items flipped right / wrong, exact binomial):

| | gained | lost | p |
|---|---|---|---|
| ARC | 5 | 13 | 0.096 |
| TruthfulQA | 32 | 13 | **0.007** |
| MMLU | 26 | 15 | 0.117 |
| HellaSwag | 10 | 10 | 1.0 |

The adapter moves B2 in the same direction it moves the stock model: up on TruthfulQA.

### Merges that didn't make it

The plan was a "DragonBophades" multi-model merge. Base models tested alongside B2 were
Ornith-1.5, MiMo-V2.6-Distill, Qwen3.8-9B-Distill and Astrea-R8-Chat (best non-B2 donor: 56.2 / 35.0 / 40.8 / 78.2).

- **model_stock (equal weight), with or without B2:** about 53.5 / 34.4 / 39.7 / 76.1. These merges land back near the base. With donors this
  different, model_stock's interpolation collapses toward the base.
- **TIES, B2 0.55 / Astrea 0.35 / Bophades 0.3:** 59.9 / 36.6 / 41.5 / 77.5. This is B2 within noise on every
  benchmark. Beating the third-party donors isn't enough of a reason to ship a model that only ties B2.

Applying the adapter directly to B2 was the only build that moved anything.

## Notes

- **Identity:** it calls itself a model made by Schneewolf Labs, which it inherits from B2. If you
  need another identity, set one in the system prompt.
- **Thinking:** trained with thinking off. Thinking-on behaviour comes from B2 and is not evaluated here.
- **Not evaluated:** generation-based benchmarks and agentic / tool-use behaviour. B2 was built for
  the [egirl](https://github.com/Schneewolf-Labs/egirl) harness, and this tune has not been re-checked there.
- Vision tower and MTP head are B2's, untouched.