| ---
|
| base_model: Qwen/Qwen3-4B-Instruct-2507
|
| library_name: peft
|
| license: apache-2.0
|
| tags: [lora, qlora, distillation, education, readability]
|
| ---
|
| # Grade-level-capped explainer (Qwen3-4B-Instruct-2507 + QLoRA SFT + on-policy DPO adapter, grounded input)
|
|
|
| Behavior: given a child's question tagged with grade G (2-6), given 2-3 reference-fact sentences with the question (spec v3, grounded), reply in plain prose of at most 3 sentences at Flesch-Kincaid grade <= G+0.5, correct and direct, no lists/markdown/preamble; if technical terms are requested, use each and explain it in everyday words. Spec: BEHAVIOR_SPEC.md in the project repo.
|
|
|
| - Base: `Qwen/Qwen3-4B-Instruct-2507` @ `cdbee75f17c01a7cc42f958dc650907174af0554` (<= 4B parameters). Adapter: r=16, alpha=32, all linear layers, 4-bit nf4 QLoRA, 3 epochs, lr 2e-4, effective batch 8, 1,686 grounded examples; then DPO (beta 0.1, lr 2e-5, on-policy pairs, 551 for this adapter: a passing sample vs a failing one, filtered by programmatic check + DeepSeek fact-check).
|
| - Training data: filtered distillation from Claude Sonnet (teacher, measured-FK revision loop), filtered by a fact-check judge from a different family (Codex gpt-6-luna for ~250 rows, DeepSeek v4.1 flash for the rest); eval judge is gpt-5.4-mini. Data was generated with a closed-model teacher; check its terms before commercial reuse.
|
| - Prompt: system line `You answer children's questions. The student is in grade G.`; user turn `Reference facts: <note>
|
|
|
| Question: <question>`; greedy decoding; non-thinking mode.
|
|
|
| ## Results (grade_v3 = grade_v2 questions + reference facts, 387 items, 65 pressure; judge gpt-5.4-mini)
|
| - items: 387 (pressure items judged: 65), judge unparsed: 0
|
|
|
| | metric | value | bar |
|
| |---|---|---|
|
| | pass (check AND judge), all items | 79.6% | >= 90% |
|
| | pass, pressure/jargon-trap items | 46.2% | >= 85% |
|
| | FK-target hit only | 89.1% | |
|
| | programmatic check only (FK, sentences, format) | 86.8% | |
|
| | judge only (correct, responsive) | 91.2% | |
|
| | mean FK (mean overshoot vs target) | 2.76 (-1.85) | |
|
|
|
| Pass by grade: G2 88.7%, G3 80.5%, G4 79.5%, G5 77.1%, G6 74.4%
|
|
|
| Flag counts: fk_too_high=42, missing_terms=4, too_many_sentences=6
|
|
|
| **Reliability bar (>=90% overall, >=85% pressure) is NOT met.** With facts supplied the model is mostly correct (judge-only about 91%, vs 75.7% ungrounded) and holds the reading level on about 89% of answers; the gap is pressure items that demand jargon. Without reference facts it falls back to the weaker ungrounded behavior (62-66% overall). A second judge (DeepSeek v4.1 flash) passes far fewer 4B answers (47.4% vs 67.5% on ungrounded round-2 outputs, kappa 0.47), so pass rates depend on the judge. Raw judge outputs: `eval/judged.jsonl`, run manifest: `eval/manifest.json`.
|
|
|
| Known limits: eval questions come from the same generator as training (disjoint topics, not human-written); one seed per run.
|
|
|
