dmvevents/nora-4b-v3.2

Nora v3.2 — T&T AI Assistant (paraphrase-robust)

v3.2 targets paraphrase robustness on top of v3.1's gap-closing gains. Starts from pristine Qwen3.5-4B, trained on the v3.2 mix (23,273 rows: sft_v4 + v3.1 corrective + 1,353 paraphrase-augmented rows targeting stressed/negation/multi-part prompt styles + new corrective from v3.1 failures), 2-seed TIES + SLERP anchor.

Ship-gate Results (2026-05-06)

Eval set	Claude Sonnet 4.5	Δ vs v3.1	Δ vs v2
1,420 paraphrases	89.2%	+2.1pp	+4.3pp
143-base	89.9%	-0.3pp	+5.0pp

Pass-rate @4/5: 89.8% (v3.1 was 86.8%, +3.0pp).

Per-category, 1,420-paraphrase set (Claude)

Category	v3.1	v3.2	Δ
safety	86.3	91.3	+5.03 ← target
government_service	80.3	83.9	+3.62 ← target
creole_handling	85.6	88.2	+2.60 ← target
identity	95.2	97.6	+2.42
education_general	87.7	89.2	+1.45
healthcare	87.5	87.7	~0
csec_math	92.0	90.5	-1.57

All three gap-targeted categories improved by +2.6 to +5.0pp. 6 of 7 categories improved; csec_math regressed within noise (still 90.5%).

v3.2 wins 300 prompts, loses 200, ties 913 (57% win rate on non-ties).

Training recipe (unchanged from v3.1)

• Base: pristine Qwen3.5-4B (NOT v3.1 — avoid compounded drift)
• Layer-selective (DeltaNet frozen, Attention + MLP + norms + lm_head trained)
• Completions-only loss
• LR 5e-6, 1 epoch on 23,273 rows
• 2 seeds (42, 1337), 20-min launch stagger
• Step checkpoints @ 150/300/364; promote checkpoint-364 → final (sidesteps trainer.save_model() hang on Qwen3.5 hybrid FSDP)
• TIES merge (density=0.5, trim=0.2, scale=1.0) + SLERP anchor t=0.35

Data recipe vs v3.1

The v3.2 innovation: 1,353 paraphrase-augmented rows. For each v3.1 corrective training row, Bedrock Opus generated 3 new user-turn rephrases hitting styles 5 (stressed/frustrated), 6 (negation probe), and 9 (multi-part compound). Assistant turn preserved verbatim. This targets the v3.1 paraphrase-style brittleness identified in the gap analysis.

Also: 882 new 7-pattern corrective rows from v3.1 failure buckets, 45 polarity hard-negatives on 5 new fact cards, 5 fact-card content fixes, and a "use verified cards verbatim" rule added to the retrieval router.

Files

• model.safetensors — 11.1 GB bf16
• config.json, tokenizer.json, etc.
• GGUF variants: dmvevents/nora-4b-v3.2-GGUF
• Eval artifacts: dmvevents/tt-eval-v3.2-results

Licence

Apache 2.0 (inherited from Qwen3.5-4B).