edgeimci-4b-alpha-second-gguf

EdgeIMCI 4B-alpha-second Q4_K_M

An EdgeIMCI research fine-tune of Qwen/Qwen3-4B, trained on the enriched 2,258-row synthetic multitask release and quantized to Q4_K_M. Previously called "Enriched 4B"; this is not 4B-alpha-first or the historical untuned 4B hardware-feasibility model.

Identity

• GGUF: EdgeIMCI-4B-alpha-second-Q4_K_M.gguf
• Bytes: 2497280288
• SHA-256: a4a8c5bb2d9f3401defa1cb8ea007812d5916c0242d8306a1c7ed6322d550919
• Base revision: 1cfa9a7208912126459214e8b04321603b3df60c
• Training run: multitask2258-20260922-v1--qwen4-e2-lr3-s20260824
• Original artifact inventory SHA-256: 7f53d7a0eac4938194fdbc977f64bd0b17bbb8fc6131f3573fcc790edfbca1a0
• Conversion: llama.cpp aedb2a5e9ca3d4064148bbb919e0ddc0c1b70ab3, merged checkpoint -> F16 GGUF -> Q4_K_M.

Fine-Tuning Provenance

LoRA SFT: rank 16, alpha 32, dropout 0.05, 2 epochs, learning rate 0.0003, seed 20260824, effective batch size 16, BF16 computation, assistant-only loss, maximum sequence length 3072, thinking disabled. This is not QLoRA. Targets: attention q/k/v/o and MLP gate/up/down projections. The terminal epoch-2 adapter was merged into the base checkpoint before GGUF conversion.

Dataset: project-authored beta0_1k_multitask_enriched_2258_v1.

Release manifest SHA-256: 0820f092ed585a46182e075477f9246917629061b03a3f57d92cb1087e303706. The enriched dataset itself is not distributed here. Public dataset availability, consolidated source-license review, and independent clinical/manual review remain pending. Apache-2.0 base/model provenance is not a license claim for every dataset source. Original adapter files are provided under adapter/; retained training and conversion evidence is under provenance/.

Evaluation Limits

On the unsealed 139-row training-validation partition the merged checkpoint had 139 JSON-valid, 135 schema-valid, 131 whole-prediction-exact, and 132 route-correct outputs. It recorded one unsafe admission out of 20 and two urgent misses out of four. These tiny safety subsets do not establish clinical reliability.

The final Q4_K_M and F16 control each passed two public extraction smoke prompts (exact JSON target and downstream deterministic state). Full quantization-drift evaluation, a matched base-versus-fine-tuned comparison, free-form qualification, and Ubuntu/ADTC laptop profiling remain outstanding. No sealed TEST data was used.

Prompt Contract

Structured extraction uses this system instruction:

When and only when the exact EdgeIMCI extraction-v2 wrapper is present, return one JSON object matching the pinned extraction-routing schema. Return the existing model-facing encounter object for recoverable encounter findings. Return only {"route":"INFORMATIONAL_NON_ENCOUNTER"} for general EdgeIMCI or project questions, {"route":"OUT_OF_SCOPE"} for unsupported requests, {"route":"OUT_OF_SCOPE_URGENT"} for explicit urgent unsupported findings, or {"route":"UNCERTAIN"} when the input cannot be interpreted safely. Never invent encounter observations to avoid a route, and never emit classifications, referrals, treatments, rule IDs, markdown, or prose.

User content is wrapped internally, without changing its observations:

<EDGEIMCI_EXTRACT_V2>
{worker_findings}
</EDGEIMCI_EXTRACT_V2>

Use the Qwen3 chat template with thinking disabled, deterministic decoding, and up to 1200 output tokens for extraction. Unknown observations remain null; never infer unmentioned findings as negative. Route objects are not encounter objects and must not be passed to the clinical engine.

Free-form assessment does not use the extraction wrapper or extraction-only system instruction. Its outputs require separate evaluation and must not be treated as deterministic clinical decisions.

Intended Use and Safety

Research and competition evaluation only. Not a medical device, not clinically validated, and not authorized for autonomous diagnosis or treatment. The intended application validates extraction JSON, requires worker review, and delegates completeness, classifications, management, and rendering to deterministic code. The model can omit findings, misread negation, hallucinate, or produce invalid output. Publishing these weights does not confer clinical approval.