Downloads · 30 days
0
think-in-silence
rajat5039/think-in-silence
think-in-silence is a question answering model from rajat5039. Use it when the input is a question plus a passage. It is set up for transformers. The card lists the license as mit.
A model that reasons in pure latent space. No chain-of-thought tokens. No RL. No reasoning traces.
Access
Public
Updated Apr 5, 2026
Repo size
1.6 GB
Likes
2
Public
Repo makeup
Click a slice to open those files.
.pt1.6 GB · 100%
At a glance
- Task
- Question Answering
- Library
- transformers
- License
- mit
- Access
- Public
- Created
- Apr 5, 2026
- Updated
- Apr 5, 2026
- SHA
- fd5f035f
Try a prompt
Model card
From the Hugging Face model README
think-in-silence
A model that reasons in pure latent space. No chain-of-thought tokens. No RL. No reasoning traces.
Results
| Thinking Steps (K) | R@1 | BLEU | ROUGE-1 |
|---|---|---|---|
| K=0 (no thinking) | 0.002 | 0.000 | 0.000 |
| K=1 | 0.064 | 0.000 | 0.028 |
| K=2 | 0.256 | 0.009 | 0.084 |
| K=4 | 0.504 | 0.044 | 0.218 |
| K=8 | 0.474 | 0.231 | 0.594 |
| K=16 | 0.406 | 0.185 | 0.542 |
Usage
from src.models.lc_thought import LCThought
from src.utils.checkpoint import load_checkpoint
import torch, yaml
from types import SimpleNamespace
# Load config and model
cfg = yaml.safe_load(open("configs/decoder.yaml"))
model = LCThought(cfg)
load_checkpoint("checkpoints/stage3/step_0050000.pt", model)
model.eval()
# Inference
answers = model.generate(q_ids, q_mask, n_steps=8)
License
MIT EOF
- Task
- Question Answering
- Library
- transformers
- License
- mit
- Languages
- en
- Created
- Apr 5, 2026
- Updated
- Apr 5, 2026