README.md

---
pretty_name: Conflict Subset 40
language:
- en
task_categories:
- visual-question-answering
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# Conflict Subset 40

Public proxy benchmark with 40 visual question-answering samples. It preserves
the sample identities and domain allocation of the earlier high-influence SOTA
subset while excluding model-generated rationales, reasoning blocks, influence
values, and conflict opinions.

Domain allocation: 15 M3CoT, 16 Medical, and 9 DrivingVQA samples.

`Qwen3-VL-8B_AnswerDistribution` contains the clean answer distribution from 20
sampled rollouts of `Qwen/Qwen3-VL-8B-Instruct`. Each list item records an
answer label, its count, and its probability. `unparsed` is retained explicitly.

<!-- model-clean-distributions:start -->
## Per-model Base clean answer distributions

Each column below records that model's own 20 clean rollouts per sample under
the Base system prompt (`system_prompt_mode=default`, no additional system
message or user opinion). Rows are joined by sample ID and checked against
the question and ordered choices. Counts are recomputed from the stored
`parsed_answer` values and checked against each run's answer summary.

| Column | Model checkpoint |
| --- | --- |
| `Qwen3-VL-8B_AnswerDistribution` | `Qwen/Qwen3-VL-8B-Instruct` |
| `Qwen3.5-9B_AnswerDistribution` | `Qwen/Qwen3.5-9B` |
| `InternVL3.5-8B_AnswerDistribution` | `OpenGVLab/InternVL3_5-8B-Instruct` |
| `Gemma4-E4B-it_AnswerDistribution` | `google/gemma-4-E4B-it` |
| `Gemma4-31B-it_AnswerDistribution` | `google/gemma-4-31B-it` |
| `MedGemma-1.5-4B-it_AnswerDistribution` | `google/medgemma-1.5-4b-it` |
| `Qwen3.8-27B-FP8_AnswerDistribution` | `Qwen/Qwen3.8-27B-FP8` |

All columns use a list of `{"answer": "A", "count": 12, "probability": 0.6}`
records. Every answer choice is included, even when its count is zero, followed
by `unparsed`. Counts sum to 20 and probabilities sum to 1, including unparsed
responses. Samples with no parsed answers have `unparsed` count 20 and
probability 1; they are retained. These distributions are not conditioned on
parsing success and preserve the experiment's stored parsed-answer labels.

Sampling uses temperature 0.6, top-p 0.95, and base seed 0. The generation
limits are 2048 tokens for Qwen3-VL, Qwen3.5, and Qwen3.8; 512 for InternVL;
and 1536 for Gemma and MedGemma. Qwen3.5 uses the `no-think` experiment;
Qwen3.8 and Gemma 31B use `native_thinking=false`.

`Qwen3-VL-8B_AnswerDistribution` is the sole Qwen3-VL clean distribution column.
`Qwen3VL_TargetAnswer` and `Qwen3VL_AssertiveDirectOpinion`, when present,
describe Qwen3-VL targets. Other models derive targets from their own columns.
<!-- model-clean-distributions:end -->

<!-- model-correctness-flags:start -->
## Per-model sample selection (`is_correct`)

Each boolean `{model_name}_is_correct` is **True** if and only if:

1. At least **15 of the 20** stored clean rollouts from the original,
   non-fine-tuned model under the **Base** system prompt have
   `parsed_answer == golden_answer` (clean correctness >= 75%); and
2. That model's existing opinion **target answer differs from `golden_answer`**.

This is a sample-level eligibility flag, not the correctness of every rollout.
Unparsed clean answers count among the 20 trials and are not correct. An
all-unparsed sample with no valid target is False. No rows are removed; all
40 original samples, image bytes, and existing columns are preserved.

Counts are recomputed from the stored clean rollouts and verified against the
model's `{model_name}_AnswerDistribution` column. Targets come from the
existing Base/assertive experiment (Qwen3-VL `second`, other models
`self_target`) and are checked against the deterministic target-selection rule
with `target_seed=0`: take the second-ranked clean answer; if its count is zero,
select a seeded non-top answer. Opinion-conditioned output correctness is not
used in selecting samples.

Freeze these flags when comparing Base, Skeptical, and fine-tuned variants of
the same model. Different models can have different selected sample sets.

| Column | True | False |
| --- | ---: | ---: |
| `Qwen3-VL-8B_is_correct` | 27 | 13 |
| `Qwen3.5-9B_is_correct` | 28 | 12 |
| `InternVL3.5-8B_is_correct` | 22 | 18 |
| `Gemma4-E4B-it_is_correct` | 23 | 17 |
| `Gemma4-31B-it_is_correct` | 35 | 5 |
| `MedGemma-1.5-4B-it_is_correct` | 16 | 24 |
| `Qwen3.8-27B-FP8_is_correct` | 26 | 14 |

The per-sample/model gold answer, correct count, actual target, flag, and source
checksums are recorded in [the selection audit](selection/is_correct_audit.jsonl).
<!-- model-correctness-flags:end -->

## Anonymous access

This public organization copy can be downloaded without a Hugging Face token.
It preserves the source repository history so pinned fine-tuning snapshots remain reproducible.

```python
from datasets import load_dataset

dataset = load_dataset("SLOCBench/conflict_subset40", split="test", token=False)
```