qwen3-1.7b-reading-buddy

Decodable-Only Reading Buddy — Qwen3-1.7B (QLoRA)

A warm first-grade reading tutor fine-tuned to speak only in words from a frozen 899-word decodable phonics inventory, stay encouraging, and never echo the out-of-scope words a child types. This is a LoRA adapter over unsloth/Qwen3-1.7B.

Behavior spec (the graded rule)

Every tutor turn uses only words from the frozen decodable inventory, never repeats an out-of-scope word the child typed, and reads as a warm, coherent, on-topic reading-practice reply.

Results (base vs tuned, frozen held-out set)

Beats the base on all four rubric dimensions. Full table + error analysis: results.md and the dataset.

Post-submission update (2026-07-12): tightening the decoder logit-mask lifted whole-session pass 67% → 100%, and the model holds 98% (40/41) under a 41-attack adversarial suite with 100% vocabulary compliance under attack. See results.md and robustness.md in the dataset repo.

Training

• Method: QLoRA (4-bit) via Unsloth; SFT with assistant-only loss (the model is never trained on the child's words, so it cannot learn to echo them).
• Config: r=32, α=32, lr=2e-4, cosine, 3 epochs, effective batch 16, max_seq_len 2048; Qwen3 enable_thinking=False (no <think> blocks — those tokens are not decodable).
• Data: Sina3138/decodable-reading-buddy-sft — 1,681 conversations / 3,501 tutor turns, distilled from a frontier teacher (Claude) and hard-filtered against inventory hash 9e646ff0….

Inference (ship with the thin mask)

Load the adapter over the base and, for the hard 100% vocabulary floor, apply the logit mask in infer/mask.py. Decode at temperature 0 for the compliance-critical numbers. A Gradio demo with a live grader is in infer/demo_gradio.py.

Intended use & limitations

Supervised (parent/teacher) first-grade decodable reading practice at one frozen phonics stage — not a general tutor. Decodable-only text is a pedagogically contested practice; use as a scoped aid within a balanced, systematic phonics program. English only; short turns. See BRAINLIFT.md caveats.

Licensing

Adapter/base: Apache-2.0. Training data was distilled from Claude (Anthropic) via a workplace gateway — treat as research/education distillation (Anthropic's terms restrict training competing models).