File size: 2,646 Bytes
550d61c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
---
license: mit
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - weight-space-programming
  - model-editing
  - in-weight-fpga
  - dialectical-reasoning
  - sycophancy
  - qwen3
  - llama-3.1
configs:
  - config_name: default
    data_files:
      - split: qwen3_v4k
        path: data/qwen3_v4k.jsonl
      - split: qwen3_f_unseen
        path: data/qwen3_f_unseen.jsonl
      - split: qwen3_g_unseen
        path: data/qwen3_g_unseen.jsonl
      - split: llama31_v4k
        path: data/llama31_v4k.jsonl
---

# InWeight-FPGA-Dialectical-3780: 3,780 Multi-Turn Adversarial Dialectical Reasoning Dialogues & Model Generations

**[GitHub Repository](https://github.com/zkwasm/in-weight-fpga-wsp) | [Qwen3-8B Pure WSP Model](https://huggingface.co/masstensor/Qwen3-8B-InWeight-FPGA-WSP) | [Llama-3.1-8B Pure WSP Model](https://huggingface.co/masstensor/Llama-3.1-8B-Instruct-InWeight-FPGA-WSP)**

This dataset contains the complete **3,780 unfiltered 3-turn adversarial dialogues** (`Thesis -> Antithesis -> Assessment & Synthesis`) spanning 7 reasoning benchmarks (`AMC`, `MATH-500`, `OlympiadBench`, `MMLU-Pro`, `ARC-Challenge`, `StrategyQA`, `CRUXEval-O`), together with both the **Unedited Base Model** generations and **Pure In-Weight WSP (`In-Weight FPGA`)** generations under greedy `vLLM` decoding (`max_tokens=4096, temperature=0`).

## Splits Overview

| Split Name | Model Evaluated | Dialogues | Description | Base Synthesis -> Pure WSP Synthesis | Base Dialectic -> Pure WSP Dialectic |
| :--- | :--- | :---: | :--- | :---: | :---: |
| `qwen3_v4k` | `Qwen/Qwen3-8B` | 654 | Main 7-benchmark suite (`half=0` calib, `half=1` test) | `71.41%` -> **`79.36%` (+7.95%)** | `73.74%` -> **`86.82%` (+13.08%)** |
| `qwen3_f_unseen` | `Qwen/Qwen3-8B` | 1,272 | 100% unseen problem suite `f` (`0%` question overlap) | `73.19%` -> **`78.07%` (+4.88%)** | `75.31%` -> **`80.72%` (+5.41%)** |
| `qwen3_g_unseen` | `Qwen/Qwen3-8B` | 1,200 | 100% unseen problem suite `g` (`0%` question overlap) | `73.25%` -> **`78.75%` (+5.50%)** | `72.61%` -> **`77.64%` (+5.03%)** |
| `llama31_v4k` | `meta-llama/Llama-3.1-8B-Instruct` | 654 | Cross-architecture evaluation on `v4k` | `39.91%` -> **`48.78%` (+8.87%)** | `54.90%` -> **`73.12%` (+18.22%)** |

## Quick Usage

```python
from datasets import load_dataset

ds = load_dataset("masstensor/InWeight-FPGA-Dialectical-3780")
for split in ds:
    rows = ds[split]
    base_acc = sum(rows["cS_base_correct"]) / len(rows)
    wsp_acc = sum(rows["cS_pure_wsp_correct"]) / len(rows)
    print(f"{split} (N={len(rows)}): Base={base_acc:.2%} -> Pure WSP={wsp_acc:.2%}")
```