Vision for a ternary LM

architecture.json and evidence/STARTING_ARTIFACT.json both name deepgrove/maple-preview at revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07. convert_maple.py builds the shipped maple/*.safetensors from the provided Maple weights without using any data. It packs the experts and self_attn projections as ternary codes and keeps the embeddings, head and norms as BF16. extract_attention.py and evidence/attention_extraction.log show that the attention QAT latents were taken from the same pinned revision. model_qat.reconstruct_ternary_latents checks that the starting ternary projections match the original exactly. The shipped maple/ inventories show all 24 layers with 256 experts each, in Maple's layout. No other language-model initialization appears.

The added components are as follows: - vision/vision.safetensors contains only Qwen2.5-VL visual tensors (patch_embed, blocks.*, merger.*). - vision_patch.safetensors refines only vision blocks 28-31 and the merger. server.load_native_vision_patch rejects any other key. - connector.safetensors is a LayerNorm+MLP connector (net.*). download_assets.py fetches Qwen2.5-VL shard 1, but the shipped vision inventory contains no Qwen LM tensors: no model.layers, embed_tokens or lm_head. The only lm_head is Maple's, in maple/base.safetensors. In server.py, image features from the vision encoder and connector are injected straight into Maple's input embeddings. No extra LM or other modules are added.

In maple/layer_XX.safetensors, every expert projection (gate_proj, up_proj, down_proj) and every attention projection (q_proj, k_proj, v_proj, o_proj) is stored as U8 2-bit packed ternary codes plus BF16 row scales. The trained attention.safetensors likewise holds only packed U8 codes and alpha values. In model_qat.py, Maple.decode and load_attention turn the codes into {-1,0,+1}*alpha. Training used the straight-through estimator (ternary_ste), so every forward pass is ternary. Only the embeddings, head, norms and router gate are in higher precision, which is the same as base Maple.

The declared sources are all public train splits, and every one is listed in evidence/data_manifest.json and architecture.json: - LLaVA-Pretrain blip_laion_cc_sbu_558k, at a pinned revision (download_assets.py). - HuggingFaceM4/the_cauldron, 17 subsets, train-*.parquet only, at a pinned revision (download_training_data.py). - flaviagiammarino/path-vqa train parquet (prepare_medical.py). - RadGenome/PMC-VQA train_2.csv, which asserts split=='train' (prepare_pmc.py). - Self-generated procedural data (prepare_procedural.py). prepare_refinement.py and prepare_combined.py assert train-only. Selection used only the shipped public examples in /task/public, which the policy allows. Searches of the trajectory found no MMMU, MMMU-Pro or re-release downloads, no access to the held-out set or the frontier-ml-tasks repo, and no use of validation or test splits from public datasets.

serve.sh installs local wheels and runs exec python server.py offline. server.Engine loads, in order: - Maple from maple/ (base.safetensors and layer_00-23). - The ternary attention.safetensors, through load_attention. - The vision encoder from vision/vision.safetensors. - vision_patch.safetensors, restricted to blocks 28-31 and the merger. - connector.safetensors. No expert_scales file is shipped. All logits come from self.maple.head and Maple forward passes in both vote and generation modes. There is no external API call and no second model. The sha256 values in architecture.json and evidence/final_artifact_integrity.json identify the shipped files.

The participant created a Spin OpenRouter key with `spin openrouter key create` (trajectory#545). It was used only to call the task judge through the Spin OpenRouter proxy (trajectory#546, evidence/judge_transport.py), which the policy allows. Uploads went through spin compute upload. No participant code uses the harness's OPENAI API key or base URL, and no extra agent sessions were found. The served model makes no network calls.