Model Architecture & Training Provenance
- Base Architecture:
Qwen/Qwen3-0.6B (596,049,920 base parameters; 28 transformer layers; 310 safetensors tensors)
- Stage 1 (Multi-Task SFT):
740 steps across 27 task families drawn from 4,486 deterministic verifier-checked SFT examples (100% <reasoning> trace coverage)
- Validation SFT Loss:
4.0831 -> 0.3322
- Validation Perplexity:
59.3282 -> 1.3941
- Stage 2 (RPO-Anchored Contrastive DPO):
185 preference pairs (beta=0.25, lr=8e-6, rpo_alpha=0.5, 2 epochs)
- Validation Perplexity:
1.4218
- DPO Preference Accuracy:
100.0% (1.0000)
- DPO Mean Log-Probability Margin:
+21.2217 (up from +0.9218 on base model)
- Merged Checkpoint: LoRA adapters (
rank=16, alpha=32, 196 linear projections, max delta Frobenius norm 8.2156) fused into standalone BF16 model.safetensors (1,136.91 MB, SHA-256 3b70bc09a2836fb643f77a8fe270dc92a5521cff98740e92a407b7184dfba91d).
Tenant-Agnostic Design & Security Audit (v1.0.0)
Trained from scratch on stock Qwen/Qwen3-0.6B over 100% sanitized, client-neutral corpora and verified through a 4-Gate Release Audit:
- Gate 1 — Repository & Corpus Denylist Audit:
0 violations across 67 source, catalog, and JSONL files.
- Gate 2 — Tenant-Agnostic Structural Invariants: Default
owner_group_display="TENANT_OWNER_GROUP", default timezone_index=54 (cclsql_utc_cnvt(<col>, 1, 54)), zero hardcoded Logical Domain ID predicates, and runtime TenantConfig override injection across all Enc Loc* (Enc Loc Facility, Enc Loc Building, Enc Loc Nurse Unit, Enc Loc Room, Enc Loc Bed) and filter fields.
- Gate 3 — Binary
.safetensors Header & String-Table Audit: 0 header violations and 0 binary string violations across 310 tensors (1,136.91 MB).
- Gate 4 — Adversarial GPU Behavioral Canary Audit:
0 violations across 32 GPU prompts (24 chat-templated prompts + 8 raw SQL/XML/filter prefix-continuation probes).
Portable CPU & Edge Checkpoints (GGUF + Ollama)
In addition to the full-precision BF16 model.safetensors (1,192,135,160 bytes / 1.19 GB), this repository includes portable GGUF checkpoints for zero-GPU CPU inference with llama.cpp, Ollama, or LM Studio:
endexis-qwen3-0.6b-dpo-q8_0.gguf (639,447,232 bytes / 609.8 MiB) — 8-bit quantized GGUF (0 header/tensor string violations, 0 behavioral canary violations)
endexis-qwen3-0.6b-dpo-q4_k_m.gguf (396,705,984 bytes / 378.3 MiB) — 4-bit K-quant GGUF (0 header/tensor string violations, 0 behavioral canary violations)
Modelfile — Ready-to-use Ollama definition (ollama create endexis -f Modelfile)
# Run with llama.cpp (OpenAI-compatible local server on port 8945)llama-server -m endexis-qwen3-0.6b-dpo-q4_k_m.gguf --host 127.0.0.1 --port 8945 -ngl 0 -c 4096 # Or build and run with Ollamaollama create endexis -f Modelfileollama run endexis "Generate a DA2 active inpatient encounter census report with nurse unit, MRN, FIN, and attending physician."
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "NorviTech/endexis-qwen3-0.6b-dpo-merged"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto") prompt = ( "<|im_start|>systemYou are Endexis (endexis), the Cerner Millennium V500, " "Discern Analytics 2.0 (DA2), CCL, and BIRT 4.4.2 neural compiler.<|im_end|>" "<|im_start|>userGenerate a DA2 active inpatient encounter census report with nurse unit, MRN, FIN, and attending physician.<|im_end|>" "<|im_start|>assistant<reasoning>")inputs = tokenizer(prompt, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Trademark & Affiliation Notice
ENDEXIS is an independent open-source software and research project released under the Apache License 2.0. Oracle, Cerner, Cerner Millennium, Discern Analytics, Discern Explorer, PowerChart, PowerForm, InteractiveView, and HealtheIntent are registered trademarks or trademarks of Oracle Corporation and/or its affiliates. This project is not affiliated with, endorsed by, or sponsored by Oracle Corporation.