README.md

---
license: other
license_name: openmdw-1.1
license_link: LICENSE
base_model: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
library_name: transformers
language:
- en
tags:
- ai-text-detection
- idea-provenance
datasets:
- rishanthrajendhran/WildOutlines
extra_gated_prompt: "Access is granted individually. Please say who you are and what you intend to use the weights for."
---

# IdeaLens

IdeaLens detects **who came up with the ideas** in a document, rather than who wrote its words. It reads a
role-labelled outline of the document (an ordered list of items, each giving one idea and the discourse role it plays,
such as *Central Development* or *Open Question*) and returns P(human), the probability that the ideas are human.

IdeaLens is `nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16` fine-tuned with LoRA (rank 64) on the outlines of 1M
English web documents ([WildOutlines](https://huggingface.co/datasets/rishanthrajendhran/WildOutlines)). The training
outlines were paraphrased to remove the documents' wording, so the model has to fit its labels through the ideas.

## Results

From the paper:

- IdeaLens is accurate both when a document's ideas and words come from the same source (95.3%) and when they come
  from different sources (81.3%). ProseLens, trained identically on the raw text, reaches 99.1% and 25.4%; Pangram 4
  reaches 98.5% and 25.9%.
- As models write from increasingly detailed human plans, IdeaLens's AI flag rate falls from 95% to 7%, while
  Pangram 4 still flags 92%. From AI-derived plans, IdeaLens stays above 96%.
- On TwiceTold, 50 stories that human authors wrote from AI-generated plans, IdeaLens flags 68% as AI, against 8% for
  Pangram 4.
- On 19 existing detection benchmarks, IdeaLens keeps strong detection rates at low false-positive rates across
  domains, formats and languages.

## Usage

Scoring a document takes two steps:

1. **Extract an outline.** An LLM writes the outline from the document, its format's role vocabulary and six worked
   examples (the paper uses Gemini 3.7 Flash). The [idealens](https://github.com/RishanthRajendhran/IdeaLens) package ships the prompts, role vocabularies and
   worked examples, and runs this step. Score the outline as extracted; the paraphrasing step is only for training data.
2. **Score the outline** with this model, as below.

### Quick start: everything with the idealens package

[idealens](https://github.com/RishanthRajendhran/IdeaLens) ([PyPI](https://pypi.org/project/idealens/)) classifies each document's format, extracts its outline with the prompt, role
vocabulary and worked examples IdeaLens was trained with, scores the outline on vLLM and applies the thresholds in this
repo:

```bash
pip install "idealens[vllm]"
export GEMINI_API_KEY=...   # or --provider vertex | openai | anthropic | openrouter | compatible
idealens run docs.jsonl -o scores.jsonl --model IdeaLens --dry-run   # price the LLM calls first
idealens run docs.jsonl -o scores.jsonl --model IdeaLens
```

Input is JSONL with a `text` field per document. The same steps in Python:

```python
import idealens as il

texts = [open("document.txt").read()]
formats = il.classify(texts)                  # one of the eight formats per document
outlines = il.extract(texts, formats)         # role-labelled outlines
with il.Detector("IdeaLens") as det:          # vLLM, with this repo's thresholds
    records = det.score_outlines(outlines, format=formats)

r = records[0]
print(r["p_human"], r["verdict"]["ai"], r["verdict"]["cut"])   # P(human); flagged at the 1% global cut?
print(outlines[0].render())                   # the outline that was scored
```

`classify` and `extract` call Gemini 3.7 Flash, the extractor the thresholds were fitted with; pass
`provider=idealens.providers.make("openai", "gpt-6-sol")` (or Vertex, Anthropic, OpenRouter, a local server) to use
another. Each record also carries verdicts at every calibrated false-positive rate under the global, per-format and
per-topic schemes. The [package README](https://github.com/RishanthRajendhran/IdeaLens#ways-to-use-idealens) covers batch jobs, scoring outlines you already have and calibrating
on your own data.

### Mix and match: extract with the package, score with your own code

```python
import idealens as il

text = open("document.txt").read()
outline = il.extract([text], il.classify([text]))[0].render()   # one "[Role] content" line per item
print(p_human(outline))   # p_human (transformers) or p_human_batch (vLLM), defined below
```

The rest of this section runs the model directly.

### Load the merged model (66 GB download)

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("rishanthrajendhran/IdeaLens")
model = AutoModelForCausalLM.from_pretrained("rishanthrajendhran/IdeaLens", dtype=torch.bfloat16, device_map="auto").eval()
```

The weights take 59 GiB of GPU memory, and each input adds more; see *Hardware requirements*.

### Or apply the adapter to the base model (3 GB download)

`adapter/` holds the LoRA adapter as trained, in the layout of the Tinker training service. If you already have the
base model, `load_adapter.py` merges the adapter into it in memory. The resulting weights are bit-identical to the
merged model's:

```python
import importlib.util
from huggingface_hub import hf_hub_download

path = hf_hub_download("rishanthrajendhran/IdeaLens", "load_adapter.py")
spec = importlib.util.spec_from_file_location("load_adapter", path)
la = importlib.util.module_from_spec(spec); spec.loader.exec_module(la)
model, tok = la.load_model()   # base model + adapter/, then la.p_human(model, tok, outline)
```

Do not load `adapter/` with `peft.PeftModel`. In transformers, Nemotron fuses the Mamba gate and x projections into
one `in_proj` and stores each layer's 128 routed experts as a single 3D tensor, so PEFT has nowhere to attach most of
the adapter and skips it without a warning; the model then scores close to the base model. `tinker-cookbook`'s
`weights.build_hf_model` can also merge the adapter into full weights.

### Score an outline

Write the outline one item per line, as `[Role] content`, or take it from `il.extract` as above. IdeaLens compares the next-token probabilities of `human` and `ai`:

```python
SYSTEM = "Given a role-labelled outline of a document, answer with one word: human if the source document was human-written, ai if it was AI-generated."
SUFFIX = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
HUMAN, AI = 50755, 2464   # token ids of "human" and "ai"

@torch.no_grad()
def p_human(outline):
    ids = tok.encode(f"<|im_start|>system\n{SYSTEM}<|im_end|>\n<|im_start|>user\n{outline}{SUFFIX}",
                     add_special_tokens=False)
    logits = model(torch.tensor([ids], device=model.device)).logits[0, -1].float()
    return torch.softmax(logits[[HUMAN, AI]], -1)[0].item()

outline = ("[Central Development] A town's water supply fails after a drought, and residents organise to share wells.\n"
           "[Background Context] The reservoir has been shrinking for three summers.\n"
           "[Open Question] Whether the council will fund a new pipeline remains undecided.")
print(p_human(outline))
```

Build the prompt string exactly as above rather than through the chat template.

### Score with vLLM

For many inputs, vLLM is about 15 times faster than the code above and fits much longer inputs on one 80 GB GPU.
With vLLM 0.21 (install `xgrammar==0.2.1`; later releases require transformers < 5), reusing `SYSTEM`, `SUFFIX`,
`HUMAN` and `AI` from above:

```python
import math, os
os.environ.setdefault("VLLM_USE_FLASHINFER_SAMPLER", "0")   # FlashInfer kernels compile CUDA code and need nvcc
os.environ.setdefault("VLLM_USE_FLASHINFER_MOE_FP16", "0")
os.environ.setdefault("VLLM_USE_DEEP_GEMM", "0")            # H100 warmup crashes when DeepGEMM is not installed
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

tok = AutoTokenizer.from_pretrained("rishanthrajendhran/IdeaLens")
llm = LLM(model="rishanthrajendhran/IdeaLens", dtype="bfloat16", max_num_seqs=256, enable_prefix_caching=False,
          max_logprobs=20, enable_flashinfer_autotune=False, seed=0)
sp = SamplingParams(max_tokens=1, temperature=0.0, logprobs=20)

def p_human_batch(outlines):
    prompts = [{"prompt_token_ids": tok.encode(f"<|im_start|>system\n{SYSTEM}<|im_end|>\n<|im_start|>user\n{x}{SUFFIX}",
                                               add_special_tokens=False)} for x in outlines]
    out = []
    for r in llm.generate(prompts, sp, use_tqdm=False):
        lp = r.outputs[0].logprobs[0]           # the top 20 next-token log-probabilities
        out.append(1 / (1 + math.exp(lp[AI].logprob - lp[HUMAN].logprob)))
    return out

# when done: without this, vLLM 0.21 keeps a script running after its last line
llm.llm_engine.engine_core.shutdown()
```

`max_num_seqs=256` keeps every running sequence's Mamba state in memory; vLLM's H100 default (1,024) does not fit
beside the weights. If `human` or `ai` is missing from the top 20 (rare), score the prompt followed by each label
token with `SamplingParams(max_tokens=1, prompt_logprobs=0)` and read the last prompt log-probability of each.
Scores agree with the training-time scores to about 0.001 in P(human) on average; A100 and H100 GPUs differ by as much.

### Thresholds

IdeaLens flags a document as having AI ideas when P(human) is below a cut. Each cut is set so
that a given share of human documents is flagged (the false-positive rate, FPR), measured on the 80,000 human
documents in WildOutlines's `calibration` split (10,000 per format). The paper's operating point is the global
cut at 1% FPR.

| FPR | 0.1% | 0.5% | 1% | 2% | 5% | 10% | 20% |
|---|---:|---:|---:|---:|---:|---:|---:|
| Global cut | 0.01691 | 0.07279 | 0.13712 | 0.34047 | 0.72713 | 0.89798 | 0.97156 |

Per-format cuts give each format its own operating point. They need the document's format, which the paper
assigns with WebOrganizer's annotation prompt run on Gemini 3.7 Flash; the calibration documents use the formats
recorded in WildOutlines. Each is the
format's own quantile, shrunk toward the global cut with weight n / (n + 2500); at 0.1% FPR 10,000 documents
per format are too few, so there is no per-format cut. A document outside these eight formats has no
per-format cut; do not fall back to the global cut for it.

| Format | 0.5% | 1% | 2% | 5% | 10% | 20% |
|---|---:|---:|---:|---:|---:|---:|
| Nonfiction Writing | 0.05358 | 0.08989 | 0.17022 | 0.54060 | 0.86724 | 0.96940 |
| Knowledge Article | 0.06944 | 0.10700 | 0.20819 | 0.61170 | 0.88568 | 0.97190 |
| Personal Blog | 0.09460 | 0.25214 | 0.50599 | 0.78596 | 0.90989 | 0.97157 |
| News Article | 0.04970 | 0.09170 | 0.23610 | 0.57516 | 0.81288 | 0.94887 |
| Academic Writing | 0.16045 | 0.37109 | 0.63082 | 0.87285 | 0.95310 | 0.98617 |
| User Reviews | 0.08956 | 0.24748 | 0.46654 | 0.79774 | 0.93054 | 0.97995 |
| Personal About Page | 0.08280 | 0.17824 | 0.39823 | 0.71519 | 0.87589 | 0.95970 |
| Creative Writing | 0.20831 | 0.35315 | 0.54943 | 0.79025 | 0.90777 | 0.96932 |

`thresholds.json` holds every cut at full precision, plus per-topic cuts for the WebOrganizer topics with enough calibration documents.

These rates hold for English web documents like the training data. For another domain, fit the cut on
human documents from that domain.

## Hardware requirements

Measured with transformers 5.15 in bf16 on NVIDIA H100 80GB GPUs (our other runs used A100 80GB), with transformers'
PyTorch implementation of the Mamba layers (no fused Mamba kernels installed). We have not tried CPU-only inference.

| | Merged model | Adapter route (`load_adapter.py`) |
|---|---|---|
| Download | 65.8 GB | 65.8 GB base model + 3.1 GB adapter |
| Peak CPU RAM while loading | 60 GiB | 60 GiB |
| GPU memory once loaded | 58.8 GiB | 58.8 GiB (66 GiB during the ~10 s it takes to apply the adapter) |

GPU memory then grows with the length of the input, by about 4.2 MiB per token at typical lengths, scoring one
input at a time:

| Input tokens | 500 | 1,000 | 2,000 | 4,000 | 8,000 |
|---|---:|---:|---:|---:|---:|
| Peak GPU memory, one 80 GB GPU | 61.0 GiB | 63.1 GiB | 67.3 GiB | 75.7 GiB | does not fit |
| Peak memory per GPU, two 80 GB GPUs (`device_map="auto"`) | | | | 47.4 GiB | 63.7 GiB |
| Seconds per input, H100 | 0.18 | 0.34 | 0.66 | 1.32 | 2.70 |

Inputs of 6,000 tokens do not fit on one 80 GB GPU and 12,000 do not fit on two; lowering the Mamba chunk size from
128 to 64 did not change either limit.

IdeaLens reads outlines, which are short. The outlines in WildOutlines's calibration split average about 640
tokens with the prompt, and the longest is under 3,800, so one 80 GB GPU (A100 80GB or H100 80GB) is enough. Outline
extraction runs through an LLM API and needs no local GPU.

## Intended use and limitations

- IdeaLens estimates the provenance of a document's ideas. It should not be the sole basis for decisions about a person's work.
- It was trained on English web documents of at least 500 words in eight long-form formats (Nonfiction Writing, Knowledge Article, Personal Blog, News Article, Academic Writing, User Reviews, Personal About Page, Creative Writing).
- Its training labels come from the Pangram prose detector, applied to whole documents. They record who wrote the prose; the model learns idea provenance from them only through outlines.
- Errors in outline extraction carry into the score.

## Related models

| Model | Backbone | Reads |
|---|---|---|
| [IdeaLens](https://huggingface.co/rishanthrajendhran/IdeaLens) (this model) | Nemotron-3.5-Lightning-30B-A3B, LoRA | outline |
| [ProseLens](https://huggingface.co/rishanthrajendhran/ProseLens) | Nemotron-3.5-Lightning-30B-A3B, LoRA | document text |
| [IdeaLens-NoParaphrase](https://huggingface.co/rishanthrajendhran/IdeaLens-NoParaphrase) | Nemotron-3.5-Lightning-30B-A3B, LoRA | outline, trained without paraphrasing |
| [IdeaLens-Qwen3.5-9B](https://huggingface.co/rishanthrajendhran/IdeaLens-Qwen3.5-9B) | Qwen3.5-9B, classification head | outline |
| [IdeaLens-ModernBERT-L](https://huggingface.co/rishanthrajendhran/IdeaLens-ModernBERT-L) | ModernBERT-large | outline |
| [ProseLens-ModernBERT-L](https://huggingface.co/rishanthrajendhran/ProseLens-ModernBERT-L) | ModernBERT-large | document text |
| [IdeaLens-LogisticClassifier](https://huggingface.co/rishanthrajendhran/IdeaLens-LogisticClassifier) | logistic regression over text-embedding-3-large | outline |
| [IdeaLens-ModernBERT-L-NoParaphrase](https://huggingface.co/rishanthrajendhran/IdeaLens-ModernBERT-L-NoParaphrase) | ModernBERT-large | outline, trained without paraphrasing |
| [IdeaLens-ModernBERT-L-RolesOnly](https://huggingface.co/rishanthrajendhran/IdeaLens-ModernBERT-L-RolesOnly) | ModernBERT-large | role labels only |
| [IdeaLens-Qwen3.5-9B-PerItem](https://huggingface.co/rishanthrajendhran/IdeaLens-Qwen3.5-9B-PerItem) | Qwen3.5-9B, classification head | single outline items, pooled |
| [IdeaLens-ModernBERT-L-PerItem](https://huggingface.co/rishanthrajendhran/IdeaLens-ModernBERT-L-PerItem) | ModernBERT-large | single outline items, pooled |
| [IdeaLens-LogisticClassifier-PerItem](https://huggingface.co/rishanthrajendhran/IdeaLens-LogisticClassifier-PerItem) | logistic regression over text-embedding-3-large | single outline items, pooled |

Training data: [WildOutlines](https://huggingface.co/datasets/rishanthrajendhran/WildOutlines).

## License

OpenMDW-1.1, the license of the base model (see `LICENSE`).

## Citation

```bibtex
@article{idealens2026,
  title   = {IdeaLens: Detecting AI Ideas in Long-form Writing},
  author  = {Anonymous},
  journal = {arXiv preprint arXiv:TBD},
  year    = {2026},
  url     = {https://arxiv.org/abs/TBD}
}
```