Commit History
-
docs(cortex): drop base_model key from card frontmatter
openhands +1: openhandsApplies the intent of PR #1 on top of current main. - Removes `base_model: deepseek-ai/DeepSeek-V4-Pro-0813` from the YAML frontmatter, so the Hub no longer renders the parent-model link. - Keeps the opening `---` delimiter. PR #1 had dropped it, which made `metadata_load()` return None: license, language, tags and pipeline_tag would all have disappeared from the Hub page. - Keeps `license: mit` in the frontmatter. Provenance is retained in the body: the "Provenance, license and attribution" table, the attribution statement, the BibTeX citation and LICENSE (MIT, Copyright (c) 2023 DeepSeek) are untouched, as the MIT license requires. Co-authored-by: openhands <openhands@all-hands.dev>f3389a6 -
feat(cortex): add CORTEX training pipeline and model audit
openhands +1: openhandsFrankenstein-Labs' own CORTEX weights need a training pipeline. This adds one, strictly separate from the distributed 1.65T checkpoint, which no script here reads or modifies. Pipeline: - model/cortex_model.py decoder-only Transformer (RMSNorm, RoPE, GQA, SwiGLU) - model/init.py weight init and checkpoint I/O - tokenizer/ loads the shipped tokenizer, validates vocab_size - training/train.py training loop with a dataset licence gate - datasets/ code datasets with declared source and licence - evaluation/evaluate.py loss, perplexity, generation - configs/cortex/ cortex-dev-1, ~174,803,712 params Correctness details found while building it: - safetensors refuses to save storage-sharing keys, which tied embeddings are. The tied matrix is stored once, aliased in provenance.json, and re-expanded on load, so checkpoints stay lossless. - Cross-entropy over full 129280-wide fp32 logits restarted the sandbox, so the softmax is compute…6fbe100 -
docs(cortex): trim redundant model card tags
openhands +1: openhandsDrop tags that duplicate Hub metadata or each other (cortex-ai, frankenstein-labs, reasoning, long-context, fp8). Keep base_model and license: mit because the weights are still the byte-identical upstream files, so the parent relation and the MIT attribution remain factually and legally required. Co-authored-by: openhands <openhands@all-hands.dev>e1985ae -
docs(agents): record CORTEX-centered model card rewrite
openhands +1: openhandsCo-authored-by: openhands <openhands@all-hands.dev>8fd8333 -
docs(cortex): center public presentation on CORTEX and Frankenstein-Labs
openhands +1: openhandsRewrite the model card around CORTEX as the product: identity block (Model/Developer/Focus), the ten capabilities, an international vision section and the founder's statement. The base model and its legal attribution stay in a dedicated provenance/license section instead of the opening. No technical claim was changed; parameter counts, context size, quantization and the unverified-checkpoint disclaimer are intact. Co-authored-by: openhands <openhands@all-hands.dev>f7ae039
-
feat: add founder behavioral charter
f019291 -
feat: deterministically intercept creator identity requests
ea6cffb -
feat: update creator identity attribution
43fd0d0 -
feat: add extensible orchestration foundation
91c4b21
-
feat(cortex-ai): agent engine, tool registry, OpenAI-compatible API, fine-tuning pipeline
Adds the CORTEX AI ecosystem layer on top of the base model. - cortex_ai/config.py configuration, environment overrides, validation - cortex_ai/tools/ tool registry with JSON-schema validation - cortex_ai/engine/ agent loop + DSML completion builders - cortex_ai/adapters/ unified model interface (mock and Hugging Face) - cortex_ai/api/ OpenAI-compatible HTTP API - cortex_ai/client.py dependency-free Python client - cortex_ai/training/ dataset pipeline, LoRA config, training entry point - docs/ French guides: architecture, API, fine-tuning - data/cortex_sft.jsonl 10-example French instruction dataset - tests/ 58 new tests Verified on a 4-core CPU machine with no GPU: the full agent loop (encode -> model -> parse -> tool -> reinject -> answer), the real HTTP API queried with curl and with the client, and the dataset pipeline validated against the real encoder. All 58 new tests pass. Not verified: load…c63bc31 -
docs(cortex-ai): correct provenance attribution and add French documentation
- name deepseek-ai/DeepSeek-V4-Pro-0813 as the upstream model and DeepSeek-AI as its author, replacing the incorrect Cortex-ai attribution in the provenance table - fix the copyright notice to match the retained LICENSE file (2023 DeepSeek) - replace the two dead upstream links with the working vLLM and SGLang pages - restore the real architecture name (DeepseekV4ForCausalLM / deepseek_v4) - restore the upstream BibTeX entry - rewrite the model card in French, centred on the CORTEX AI project - reword identity.py so the system prompt no longer claims the base model was created by Frankenstein-Labs; also fix accent/apostrophe matching in is_identity_question, which previously failed on 'Qui t a cree ?' - add docs/: getting started, message encoding and tool calling, distributed inference, and the real state of the vision layer The weight files are unchanged.2763f1c
-
Update README.md
Verified signature6214f10 -
Merge upstream docs commits from Frankenstein-Labs/Cortex-ai
Abdoulaye coumbassa +1: openhandsBrings in two documentation-only upstream commits: 1895cec docs(cortex-ai): emphasize project identity in model card 0afd8ee docs(cortex-ai): remove base model tree link Both touch only README.md, which the local stage-1 vision projector work does not modify, so the merge is conflict-free. The 66 DeepSeek-V4-Pro shard LFS pointers are byte-identical before and after. Co-authored-by: openhands <openhands@all-hands.dev>afc6969 -
feat(vision): trainable stage-1 vision projector with frozen encoder and LM
Abdoulaye coumbassa +1: openhandsImplements the first genuinely trainable CORTEX vision pipeline, isolated from the DeepSeek-V4-Pro checkpoint: IMAGE -> real SigLIP encoder (frozen) -> VisionProjector (trained) -> PrefixFusion -> small real causal LM (frozen) -> logits Only VisionProjector is trained. On the CPU sandbox the run converges from loss 7.5445 to 5.1408 over 60 steps on real Flickr30k image/caption pairs, with the encoder and LM parameter fingerprints unchanged at exactly 0. - cortex_ai/vision/train.py: config, freezing helpers, fused labels/mask, training loop, evidence report and separate projector checkpoint. - examples/train_vision_projector.py + configs/vision_projector.json: reproducible CPU entry point; every config field is also a CLI flag. - VisionProjector gains a two-layer 'mlp' type; 'linear' stays the default. - Fix the same BatchFeature/UserDict bug in VisionAdapter.encode_images that was fixed in native.py, and extend it with encode_pixels for preprocessed batches. - te…ba3bbc7 -
docs(cortex-ai): remove base model tree link
Frankenstein-Labs0afd8ee -
docs(cortex-ai): emphasize project identity in model card
Frankenstein-Labs1895cec
-
fix(vision): unpack processor output by keyword in native fusion
Abdoulaye coumbassa +1: openhandsTransformers image processors return BatchFeature (a UserDict, not a dict subclass), so the isinstance(..., dict) check failed and the processor output was passed positionally into the encoder's input_ids parameter. Against a real encoder this raised ValueError: You have to specify input_ids. Test against collections.abc.Mapping instead, and add a regression test using a UserDict-based processor that fails on the previous code. Also add AGENTS.md recording verified repo facts, sizing constraints, licensing corrections and the small-model testing approach. Co-authored-by: openhands <openhands@all-hands.dev>c095142 -
feat(cortex-ai): add native vision fusion prototype
Frankenstein-Labsa82e5d8 -
feat(cortex-ai): add executable vision bridge
Frankenstein-Labs66e3394 -
feat(cortex-ai): establish identity and vision architecture
Verified signaturef5eae9c -
Cortex AI: rewrite model card (audited specs, verified provenance)
Verified signaturebe2363b -
Update README.md
Verified signatureafef00e -
Update README.md
Verified signaturedf3d15e -
Rewrite README as CORTEX AI model card
Frankenstein-Labsf9c4a89 -
Cortex AI: model card with DeepSeek provenance and attribution
Verified signature41a5f51 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturee747ae5 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature2480052 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturee0039a5 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturec9a3f1c -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturef636360 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturebcf111d -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturedc05f8c -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature8fe195a -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature5552ee5 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturef2bc19d -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature1601678 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturecf27932 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature0faedbc -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature83b79ef -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature89c1d05 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature3fb48bf -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature96a73e9 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signaturec367dec -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signatureaa19b6f -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature5d8c4e4 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature5955064 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature043f160 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signatureef1c2b5 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature606f593 -
Copy files from models/deepseek-ai/DeepSeek-V4-Pro-0813
Verified signature334c75f