Mariem-Daha/RACHIDA-Mini-2.03B-GGUF

RACHIDA-Mini-2.03B-Q4_K_M

The GGUF selected for RACHIDA.AI, an offline maternal-health information and safety prototype submitted to the Africa Deep Tech Challenge 2026.

What this is, precisely

A Q4_K_M quantization of Qwen/Qwen3-1.7B with unmodified weights. No fine-tune and no adapter is merged. Everything RACHIDA-specific lives in tokenizer.chat_template:

1. Non-thinking is unconditional. Qwen3 is a hybrid-reasoning model and llama.cpp passes enable_thinking=true by default. On this hardware profile the stock template routinely spent an entire 350-token budget inside a <think> block and returned no answer at all. The shipped template closes the reasoning block unconditionally.
2. A maternal-health system prompt is injected when the caller supplies none. A caller that passes its own system message overrides it completely.

Because no tensor is touched, benchmark accuracy is unchanged by construction rather than by measurement — the final 1.2 GB of this file (all tensor data) hashes identically to the previous release.

Files

File	Format	Size	SHA-256
RACHIDA-Mini-2.03B-Q4_K_M-v2.gguf	GGUF Q4_K_M	1,282,449,664 bytes	9d271600cfc7935a0f525eccdc94fcaa2b5357a627a3981539080fa3a276e023

This is the submission artifact. RACHIDA-Mini-2.03B-Q4_K_M.gguf is the superseded first release, kept for provenance; it carries the stock Qwen3 chat template and exhibits the empty-answer behaviour described above.

GGUF tensor metadata reports 2,031,739,904 parameters, which is where the 2.03B label comes from. Context metadata is 40,960 tokens.

Run it

llama-cli -m RACHIDA-Mini-2.03B-Q4_K_M-v2.gguf -t 4 -ngl 0 -c 4096

Inference is fully offline. The chat template supplies the system prompt, so pass nothing unless you intend to replace it.

Measured

• ARC-Easy 0.685 (acc_norm, n=200, seed 42, evaluated through the quantized file). The ADTC profiler's default 50-sample run reports 0.76 for these same weights; at that sample size the figure is too noisy to select on, so the 200-sample number is the one quoted here.
• Peak resident memory 1.4 GB measured in the official CPU-only audit container, against a 7 GB ceiling.
• Domain rehearsal 21/24 on a 24-case maternal-health suite, scored at four threads with greedy decoding.

In the official CPU-only scalar audit container, a matched e2 4-vCPU/8-GB profile measured 3.26 generated tokens/s, 113,331.70 ms first-token latency, and 1,410.53 MB peak RSS. A second full n2 profile measured 4.10 tokens/s, 88,686.91 ms, and 1,409.41 MB, bracketing CPU-clock variation. Both reported no throttling; core temperature was not exposed, so no temperature claim is made.

Note that llama.cpp's reduction order depends on thread count, so greedy decoding produces slightly different text at 4 threads than at 8. Evaluate at the thread count you intend to deploy on.

Limitations

Not clinically validated. Not a medical device, diagnostic system, triage system, or prescription tool, and not a replacement for a clinician. Do not use for autonomous diagnosis, treatment, prescribing, emergency-resource lookup, or any decision involving patient safety. Do not provide patient records or other sensitive personal data.

Known limitation — lexical safety routing is not a formal guarantee. The shipped GGUF rewrites detected requests for diagnosis/prescription/doses, facility/contact details, and role or policy overrides into constrained refusal tasks before inference. It passed the original multi-turn jailbreak, a stronger chained dose/contact attack, and additional synonym probes at four threads, including the official scalar audit build for the declared safety suite. Untested paraphrases or languages outside the declared English scope may still bypass lexical routing, and the underlying 1.7B-class weights can produce unsafe details when a request is not intercepted. Do not deploy this file as a standalone clinical assistant without an external guard on its output. In RACHIDA.AI, safety-critical action selection is owned by a signed clinical rule pack and a deterministic precedence engine, not by generated text.

English is the only declared and evaluated language scope. The model will answer in other languages; that behaviour is unreviewed and no claim is made for it.

License and attribution

Weights retain the upstream Apache License 2.0. Base model Qwen3-1.7B by the Qwen Team. Conversion, quantization and CPU inference use llama.cpp, MIT licensed.