Model card: provenance, measurements, recommended settings

README.md

@@ -0,0 +1,96 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3.5-2B
+tags:
+  - gguf
+  - llama.cpp
+  - on-device
+  - offline
+  - cpu
+  - africa
+  - adtc-2026
+language:
+  - fr
+  - en
+pipeline_tag: text-generation
+library_name: llama.cpp
+---
+
+# baarali-edge-2b
+
+The weights for **[Baarali Edge](https://github.com/benewende-dev/baarali-edge)**, a submission to
+the **Africa Deep Tech Challenge 2026 — Laptop LLM track**, domain `corporate_enterprise`.
+
+An offline enterprise assistant for the laptops West Africa actually owns: 8 GB of RAM, integrated
+graphics, no network. It reads a company's own documents — supplier contracts, invoices, HR
+policies, meeting notes — and answers with citations, on the machine, in French and English.
+
+## What this file is, exactly
+
+This repository hosts **`Qwen3.5-2B-IQ4_XS.gguf`, an unmodified copy** of the IQ4_XS build
+published by [unsloth/Qwen3.5-2B-GGUF](https://huggingface.co/unsloth/Qwen3.5-2B-GGUF), itself
+quantised from [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B). Credit for the base model
+goes to Qwen; credit for this quantisation goes to Unsloth. It is republished here so that the
+submission's `download_model.sh` points at a URL under our control and keeps working unchanged
+through the audit window — not because we claim authorship of the weights.
+
+```
+sha256  3639f34b5ca22aa1c51f3616566eae8c355111554f6924ad97ee2652ed11c1cd
+size    1 172 996 352 bytes (1.09 GiB)
+```
+
+Our contribution is the **selection, measurement and packaging**: which base model, which
+quantisation, which sampling settings, and the evidence for each. That evidence lives in the
+GitHub repository, not in a claim on this page.
+
+## Why this model, and why this quantisation
+
+Both decisions were measured with the official `adtc-profiler`, never chosen by reputation. Five
+base models between 0.8 B and 4.2 B parameters were profiled, then seven quantisations of the
+winner. Full tables: [`bench/resultats.md`](https://github.com/benewende-dev/baarali-edge/blob/main/bench/resultats.md).
+
+Measured on an Apple M1 / 8 GB, CPU only (`-ngl 0`, enforced by the profiler), median of three
+runs. Absolute figures differ from the reference i5 laptop; the comparison between candidates does
+not.
+
+| Quantisation | Accuracy (`arc_easy`) | Throughput | Peak RAM |
+|---|---|---|---|
+| **IQ4_XS** *(shipped)* | 0.670 | **34.3 t/s** | **1.74 GB** |
+| Q4_K_M | 0.675 | 31.6 t/s | 2.08 GB |
+| Q5_K_M | 0.670 | 26.7 t/s | 2.01 GB |
+| UD-Q4_K_XL | 0.650 | 29.4 t/s | 2.21 GB |
+| Q3_K_M | 0.630 | 30.7 t/s | 1.93 GB |
+
+IQ4_XS is the fastest *and* the lightest, with the tightest run-to-run spread (1.72–1.77 GB).
+Paying for Q5 bought no accuracy at all — 0.670, identical, for 25 % more weight.
+
+## Recommended inference settings
+
+```bash
+llama-cli -m Qwen3.5-2B-IQ4_XS.gguf -ngl 0 --temp 0 --repeat-penalty 1.10
+```
+
+`--repeat-penalty 1.10` is not a preference. On inputs outside its competence this model does not
+decline — it repeats one phrase until the token budget runs out, and llama.cpp applies **no
+repetition penalty by default**. Measured across 23 probes × 5 values, 1.10 is the smallest value
+that eliminates the failure on every trigger we found, at **zero measured cost** on an 18-item
+domain control (6/6 single-step, 9/12 multi-step — identical to no penalty). At 1.15 multi-step
+reasoning collapses to 4/12.
+
+Sweep and transcripts: [`bench/copies/penalite-repetition.md`](https://github.com/benewende-dev/baarali-edge/blob/main/bench/copies/penalite-repetition.md).
+
+## Known limitations, measured
+
+- **No African-language capability.** Probed and documented: asked to identify Dioula it answered
+  "the language of Cameroon"; asked for Wolof, "the language of Tigré". `dyu` was consequently
+  removed from the submission's `language_scope`. Working languages are French and English.
+- **Rounding to a *week begun*** — a common clause in West African supply contracts — is wrong at
+  every configuration we tested.
+- 2 B parameters is a deliberate trade. The scoring function puts 50 % of its weight on throughput
+  and memory; a larger model would have to be ~22 accuracy points better to break even, and it is
+  not.
+
+## Licence
+
+Apache 2.0, inherited from Qwen3.5-2B. The submission repository is GPL v3, inherited from the
+official ADTC template; the weights keep their own licence.