File size: 2,295 Bytes
758fdf4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
57
58
59
60
61
62
63
64
65
66
67
68
---

license: apache-2.0
base_model: OpenGVLab/InternVL3-2B
library_name: peft
tags:
- lora
- polyvalent
- internvl3
- logical-reasoning
- deduction
- reasoning
pipeline_tag: text-generation
---


# Polyvalent Logical Reasoning Specialist Adapter (v3_sft)



PEFT LoRA adapter enhancing formal deduction, propositional logic, and puzzle-solving on the **InternVL3-2B** base model.



## Model Description

- **Adapter Name:** `v3_sft` (Reasoning Specialist)
- **Base Model:** `OpenGVLab/InternVL3-2B` (pinned at revision `899155015275a9b7338c7f4677e19c784e0e5a21`)
- **LoRA Configuration:** Rank $r=64$, $\alpha=128$, dropout $0.05$ on `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`.
- **Base Weights:** Apache-2.0

## Training Mixture Attributions
Trained on a curated multi-source reasoning mixture:
- **Dolly-15k:** Databricks (CC BY-SA 3.0)
- **ProofWriter:** Allen Institute for AI (CC BY 4.0)
- **FOLIO:** First-Order Logic dataset (CC BY-SA 4.0)
- **Self-Generated Puzzles:** Synthetic propositional logic problems.
- *Notice:* Adapter weights are released under Apache-2.0. Any downstream derivative works incorporating the training mixture data fall under CC BY-SA 4.0.

## Benchmark Performance
Evaluated against rigorous deduction benchmarks:
- **LogicBench:** Multi-step symbolic constraint satisfaction
- **LogicNLI:** Natural language inference with formal reasoning
- **Multi-LogiEval:** Multi-hop deductive reasoning suites

## How to Load

```python

import torch

from transformers import AutoModel, AutoTokenizer

from peft import PeftModel



BASE_MODEL = "OpenGVLab/InternVL3-2B"

BASE_REVISION = "899155015275a9b7338c7f4677e19c784e0e5a21"



tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, revision=BASE_REVISION, trust_remote_code=True)

base_model = AutoModel.from_pretrained(

    BASE_MODEL,

    revision=BASE_REVISION,

    torch_dtype=torch.bfloat16,

    trust_remote_code=True

).eval()



model = PeftModel.from_pretrained(base_model, "Pushkarrrr15151/PolyVL-logic-lora").eval()

```

## Citation
```bibtex

@article{polyvalent2026,

  title={Polyvalent: Single-Base Multi-Specialist Architecture with Dynamic Adapter Residency},

  author={Polyvalent Research Team},

  year={2026}

}

```