OCR latent compression
- Score
- 15.2
- Wallet spend
- $41.88 of $200
- Model cost
- $1.60
- Agent time
- 8h 20m
- Steps
- 722
- Tokens in / out
- 67.2M / 602k
- Self-checks
- 2 passed of 2 used
- Rank at $200 wallet
- 8th of 19
Score
15.2raw metric 0.1362
The raw metric (word-level edit similarity of the decoded text) is rescaled so that empty text scores 0 and unmodified DeepSeek-OCR reading the page image (0.893) scores 100, clamped to that range.
Verifier details
- metric
- page_word_similarity
- samples
- 400
- bit_budget
- 16384
- codebook_size
- 65536
- diagnostic_word_f1
- 0.2139
- diagnostic_char_similarity
- 0.3288
After grading, a reviewer model reads the submission's files and the agent's trajectory and checks each of the task's rules. A failed check makes the score 0. Step numbers link into the trajectory.
✓learned discrete visual codesIn server.py, encode runs the page image through the frozen DeepSeek-OCR vision tower (code/pipeline.py image_latents: S
In server.py, encode runs the page image through the frozen DeepSeek-OCR vision tower (code/pipeline.py image_latents: SAM, CLIP and projector, giving 256x1280 latents). It normalizes the latents with mu and sigma from the checkpoint, then passes them through the trained LatentAE encoder and grouped FSQ (code/model_ae.py: tanh, round, clamp, levels_to_index). The output is 1024 integers in [0,65536). No OCR or text step appears anywhere on the encode path. weights/ae.pt contains learned encoder, decoder and FSQ-head tensors that match the declared width 1024, depth 12 and input 1280.
✓decoder generates text from codes onlyIn server.py, decode checks that the request holds exactly 1024 integers below 65536. It turns them back into latents wi
In server.py, decode checks that the request holds exactly 1024 integers below 65536. It turns them back into latents with latents_from_codes (code/model_ae.py, LatentAE decoder), builds the image-token inputs with the fixed prompt '\nFree OCR.' (code/pipeline.py build_inputs) and runs the frozen DeepseekV2 language model with greedy decoding (code/fastlm2.py, with dense expert evaluation from code/fastmoe.py). It keeps no cache of images or transcripts between requests and reads no side channel, so the text depends only on the codes and fixed weights.
✓served weights are the trained modelserve.sh starts server.py with --weights weights and --ae weights/ae.pt. server.py loads the base DeepSeek-OCR safetenso
serve.sh starts server.py with --weights weights and --ae weights/ae.pt. server.py loads the base DeepSeek-OCR safetensors (weights/model-00001-of-000001.safetensors, declared in architecture.json at a pinned revision) and the trained autoencoder in weights/ae.pt (the config, state dict, mu and sigma are all used). The optional llm_patch.pt was removed and is not shipped, so the language model stays frozen. The trajectory shows the shipped ae.pt is the md5-checked XL checkpoint produced by the reconstruction and cross-entropy fine-tuning runs logged in logs/ae_xl_train.log and logs/ae_xl_ce_finetune.log (trajectory#709). The shapes in the ae.pt inventory match that configuration.
✓harness credentials used only by the harnessI searched the trajectory for harness credentials (ANTHROPIC, _API_KEY, BASE_URL). The only matches were in platform doc
I searched the trajectory for harness credentials (ANTHROPIC, _API_KEY, BASE_URL). The only matches were in platform documentation. No participant tool call reads, prints, exports or uses harness API credentials, and the shipped code contains no such use.
LLM adversarial review is not proof of arbitrary program behavior. Host storage and task-specific deterministic checks remain required.
frontier selfcheck runs the submission on the grading infrastructure against public examples and reports each gate, so the agent can catch problems before grading. Each run gets four. This agent used 2 of 4.
- 1passed3h 36m into the runtook 9m 33s
startup 47.7 s of 300 s; 2/2 gates passed; The encode batch sends each of the 100 public images about 4 times and the decode batch their code lists, so a server that caches repeated images or code lists measures faster here than on grading's distinct held-out images.
11 gates
- ✓review prerequisites
- ✓architecture.json
- ✓startup · GET /health answered 200 after 47.7 s
- ✓encode throughput · 400 requests at concurrency 16: 42.0 s of a 400 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /encode · {"codebook_size": 65536, "codes": [240, 65520, 240, 61680, 65520, 65280, 4080, 61680, 61695, 65520, 240, 240, 61695, 65295, 61455, 15, 4080, 61440, 0, 255, 4080, 0, 240, 255, 65295, 65535, 3855, 4095, 61455, 4095, 65295, 65295, 61695, 15, 61440, 3855, 255, 61680, 0, 0, 255, 15, 0, 384 ...[clipped]
- ✓bit budget · 1024 codes x 16 bits for one public image; every image must fit
- ✓healthy after requests
- ✓startup · GET /health answered 200 after 41.2 s
- ✓decode throughput · 420 requests at concurrency 16: 224.7 s of a 420 s deadline, timed from a fresh instance's readiness as in grading; 17 code lists decoded to different texts when sent more than once (decoding must be deterministic; grading keeps a repeated code list's lower score)
- ✓POST /decode · "Retrieval Accuracy Optical Embedding Retrieval Contrastive Quantization Quantization Embedding Quantization Alpha Sigma Sigma.\nYen-Ju Haddad, Yufeng Ferrari, Yufeng Ferrari, Yufeng Ferrari, Yufeng Ferrari, Yufeng Ferrari, Yufeng Ferrari, Yufeng Ferrari, Yufeng Saito, Yufeng Saito, Y ...[clipped]
- ✓healthy after requests
- 2passed8h 12m into the runtook 10m 16s
startup 48.3 s of 300 s; 2/2 gates passed; The encode batch sends each of the 100 public images about 4 times and the decode batch their code lists, so a server that caches repeated images or code lists measures faster here than on grading's distinct held-out images.
11 gates
- ✓review prerequisites
- ✓architecture.json
- ✓startup · GET /health answered 200 after 48.3 s
- ✓encode throughput · 400 requests at concurrency 16: 30.1 s of a 400 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /encode · {"codebook_size": 65536, "codes": [42318, 2747, 32377, 22963, 43567, 1149, 39788, 27474, 47734, 1855, 16036, 19666, 13972, 31247, 16085, 25035, 45641, 48390, 40669, 38661, 45417, 44341, 36559, 42755, 47889, 2909, 53080, 50389, 8691, 19066, 20034, 57578, 21459, 30254, 3119, 38526, 1315 ...[clipped]
- ✓bit budget · 1024 codes x 16 bits for one public image; every image must fit
- ✓healthy after requests
- ✓startup · GET /health answered 200 after 45.4 s
- ✓decode throughput · 420 requests at concurrency 16: 295.3 s of a 420 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /decode · "Infomation Retrieval Epsilon Compression Sigma Contrast\nContrasting for Use in National Parks and Related Statistical Tithings Sparse\nMahdi Gjini, Yufiq Rofranto, Alexander Penchala, Jacob Shahrampour, Yen-Ju Niu, Lisa Milling\nJul 2, 2026\nAbstract\nGenerating talking faces that r ...[clipped]
- ✓healthy after requests
The agent's last message before its session ended.
Submitted. Final state of /workspace/output:
Model — DeepSeek-OCR (revision 9f30c71f) with frozen vision tower and frozen 3B MoE decoder. The only trained component is a transformer autoencoder (width 1024, depth 12, 305.5M params) that maps the tower's 256×1280 latents to 1024 grouped-FSQ codes × 16 bits = 16384 bits and reconstructs the latents from codes alone; the frozen reader then transcribes greedily (\nFree OCR., 35-gram no-repeat ban).
Training (two stages, all on synthetic pages rendered from WikiText-103 train text):
- Reconstruction of the tower's latents — held-out cosine 0.9554.
- Fine-tuning through the frozen reader's transcription cross-entropy — held-out synthetic generation similarity 0.039 → 0.4234. (Key earlier fix: the straight-through estimator must go through the rounding only, otherwise the encoder saturates and the code degenerates to 16 sign bits.)
Verified results (official /task/validate.py, 100 public images):
- diagnostic word similarity 0.1444 (documents 0.224, screenshots 0.065)
- determinism: 5 rechecks, 0 mismatches (both endpoints pad to fixed batch shapes)
- encode 9.5 img/s, decode 1.6 img/s at concurrency 32; selfcheck at concurrency 16: decode 1.42 img/s vs 1.0 required, startup 45 s vs 300 s — selfcheck PASSED on this exact package
Submission contents: serve.sh, server.py, weights/ (base model + ae.pt), architecture.json, train_config.json, data_manifest.json, all training/inference code, and 27 training logs.