Exact Latents

*A real playground.py session: a held-out function comes back byte-identical from 42 continuous vectors, then answers a question straight from the same vectors (CPU run, waits trimmed).*

Write-up: readcc.net/posts/exactlatents

exactlatents compresses Python functions into continuous latent vectors that a fine-tuned decoder reads in place of source tokens, then reconstructs never-seen code byte-identically. The encoder, learned projector, and Qwen3-1.7B decoder expose a compact continuous interface while keeping exact source recovery as the primary metric.

577/600 code-exact and 527/600 byte-exact on 600 post-cutoff out-of-distribution functions, using about 3.7× fewer decoder context slots.

Quickstart

The released weights are hosted on Hugging Face. From the repository root:

uv sync
hf download labguy/exactlatents-qwen3-1.7b --local-dir weights/
uv run playground.py

The download supplies weights/model.safetensors and weights/projector_ema.safetensors.

Results

Evaluation	Released checkpoint	Comparison	Evidence
OOD-600, post-cutoff Python functions	527/600 byte-exact; 577/600 code-exact at 3.73× fewer decoder context slots	Same-set fine-tuned text-copy: 565/600 byte-exact; 584/600 code-exact	Results · OOD receipts · control receipts
Canary development set	30/36 byte-exact; 35/36 code-exact	Model-selection set, not the headline OOD evaluation	Results · canary receipts
Fully OOD function QA	39.0% with latents	39.6% with full text; statistically indistinguishable	Results · QA receipts
General-task retention spot-check	19/24	Stock Qwen3-1.7B: 24/24	Results · retention receipts

The reduction ratios above count decoder context slots and corresponding KV-cache entries, not stored bits. A 2,048-dimensional bf16 latent vector is roughly a 500× bit expansion over a token ID, so this is not a claim of bit-level compression. The QA result is statistical parity, not an improvement over text. All reported measurements use a single seed, one language (Python), and one base model (Qwen3-1.7B).

Repository layout

• compressor/ — model, latent projection, generation, and exactness APIs
• train/ — training entry points and configuration
• eval/ — reconstruction, QA, controls, and analysis entry points
• dataset/ — corpus construction and evaluation-set tooling
• docs/ — method, training, and result documentation
• receipts/ — published JSONL evidence grouped by evaluation
• data/ — shipped evaluation inputs
• weights/ — downloaded checkpoint and projector weights

Citation

Please cite both this repository and the Qwen3 base model:

@software{exactlatents2026,
  author = {{LabGuy94}},
  title = {exactlatents: Exact Python Reconstruction from Continuous Latents},
  year = {2026},
  url = {https://github.com/LabGuy94/exactlatents}
}

@misc{qwen3_technical_report,
  author = {{Qwen Team}},
  title = {Qwen3 Technical Report},
  year = {2025},
  eprint = {2505.09388},
  archivePrefix = {arXiv},
  url = {https://huggingface.co/Qwen/Qwen3-1.7B}
}

Acknowledgments

Trained on 8xH100 with the gracious support of givemeanode.

License

Apache License 2.0. See LICENSE.