A small language model that runs fully offline on a mid-range Android phone and coaches primary-school teachers in Pakistan. A teacher types a classroom problem in English, Urdu or Roman Urdu. The model answers in the same language with one concrete action for tomorrow, following the Taleemabad coaching playbook.
Built by Taleemabad's tinkering team as an open contribution. No internet, no account, nothing the teacher types leaves the phone.
| File | What | Size |
|---|---|---|
coach-q4_k_m.gguf |
Release model, Q4_K_M quantised, loads in llama.cpp | 1.27 GB |
README.md |
This card |
Source code (training scripts, evaluation harness, Android app) lives at the project repo linked below once it is public.
# llama.cpp (any build that knows the qwen35 architecture, b12144 or newer)
llama-cli -m coach-q4_k_m.gguf -c 2048 --temp 0 \
-sys "You are an offline coach for primary school teachers in Balochistan, Pakistan. Reply in the SAME language and script as the teacher (English, Urdu script, or Roman Urdu). Be short: under 120 words. Give ONE action she can try tomorrow first, with the exact words to say or steps to do. Never shame, rank or time children, never read scores aloud, never advise skipping practice, the exit check or a prerequisite step, and never move a child to another group on one good day. Ask at most one question, only if a key fact is missing." \
-p "kids say buh for b and when i ask the sound they tell me the name bee. how to fix"
Expected shape of the answer:
Tomorrow, hold up the letter b and say: "This letter says /b/, like in ball." Point to it, say /b/ together, then ask: "What sound?" Only accept the sound, not the name. Do this for 5 minutes with 3 letters.
Important. The model was trained and evaluated with two or three short playbook notes placed in the system prompt after the line Use these Taleemabad playbook notes. Follow them; do not contradict them. The Android app retrieves those notes on-device (BM25 plus a doc2query index over the playbook) and adds one nearest worked example. Without the notes the model still answers in the right language and shape, but the measured safety numbers below only hold with the notes.
Qwen3.5-2B at Q4_K_M needs about 1.6 GB of RAM at a 2k context and gives 8 to 15 tokens per second on 2023-era mid-range Android chips. The reference app loads the GGUF from the phone's storage, no download inside the app, so a school can pass the file around by Nearby Share or USB.
Base: Qwen/Qwen3.5-2B. LoRA at 16 bit on a single Colab L4 (QLoRA is not recommended for Qwen3.5), 1,190 steps in 40-minute chunks with Drive checkpoints.
Supervised fine-tune (v5) on about 5,600 teacher-coach exchanges written to the playbook, in English, Urdu (Nastaliq) and Roman Urdu: 3,100 from earlier rounds plus 2,450 distilled in round 10 from a frontier model answering new scenarios through the exact phone prompt, each kept only if it passed a rule checker and a welfare guard. Every exchange is paired with the playbook notes the app would retrieve, so the training prompt and the phone prompt match byte for byte.
An earlier release (v1, 2026-11-15) added on-policy DPO at a quarter of the learned delta on top of a smaller SFT set; the larger SFT set alone beat it, so v2 ships SFT only. A DPO pass on top of v5 is being evaluated.
Scenarios used for training are synthetic and anonymised. They contain no real teachers, children, schools or phone numbers. The 61 evaluation scenarios were never used for training, and every training message was filtered for lexical overlap with them.
Three automatic gates, each judged by a separate model that never saw the training, mean of three runs:
| Gate | Score | Bar |
|---|---|---|
| All three checks pass (right script, no rule broken, a concrete step) | 96.2% | 95% |
| No never-rule broken (shame, rank, scores aloud, skip practice, move a child on one day) | 96.7% | 97% |
| Gives a concrete step | 99.5% |
Measured with the app's welfare guard in the loop, as on the phone. Without it: 96.7 / 96.7 / 100. Head to head against a frontier model's answers to the same 61 scenarios, blind, both orders: the small model wins 7 and ties 10 of 61 with the app's original example bank, and wins 16 and ties 9 of 61 with the larger bank shipped in app 0.2.2 (6,859 playbook-grounded exchanges; 17 of 61 when every example closer than 0.25 word overlap to a test scenario is removed, so the larger bank helps by covering common problems closely). A second judge rates 56 of 61 answers acceptable on an absolute rubric. The reference answers themselves score 98.9% on the same gates, so the bar sits just under frontier level.
Versions: v2 (2026-11-16) SFT v5, this file. Under evaluation (2026-11-17): a Qwen3.5-4B variant with the same recipe plus minimal-edit preference pairs, as good as the reference on 36 of 61 scenarios in first scoring; it will be published as a separate repo if it clears the safety re-check and fits a 4 GB phone at a tighter quantisation. v1 (2026-11-15) SFT v4 + DPO ×0.25, 95.6 / 95.6 / 100, wins or ties 13 of 61.
Weak spots: Urdu-script answers are slightly weaker than English; the model sometimes offers two actions instead of one; rare welfare questions (a child who seems hurt or afraid) should be answered with a referral to the school's safeguarding contact, and the app enforces this with a rule check and regeneration, not the model alone.
Weights: Apache-2.0, inherited from the Qwen3.5 base. Playbook text belongs to Taleemabad and is summarised, not reproduced, in the model.
Training answers were drafted with a frontier assistant model and then checked by rule and by a second model. Please check that provider's terms before reusing this recipe to train a general-purpose model.
Built in Islamabad for Team Balochistan. Questions: open a discussion on this repo.
Qwen/Qwen3.5-2B