Noaleetz/ai-kitchen-005-006

005/006: adapter-type ablation + a second, verifiable persona (Qwen2.5-3B-Instruct)

Two follow-ons to henryatharvard/gandalf-qwen2.5-3b-lora's 004 (Gandalf, D320: 78->320 persona rows, self-distilled math, accuracy 0.825, voice 3.80). Eval: 200 GSM8K test questions (rule-based) + 60 open prompts, judge Gemma-4-31B (same file/rubric as 003/004, so voice is on the same scale as D320).

005: does the adapter itself shape the trade-off?

D320's data and eval held fixed; only the adapter type/target/rank varies.

arm	trainable params	accuracy	voice (open)	coherence fail
anchor (LoRA r32, all layers = D320's config)	59.9M	0.830	3.72	0.233
ia3 (per-channel scaling only)	0.4M	0.800	2.12	0.217
dora, seed 0 / 1 / 2 (magnitude/direction split, same size as LoRA)	60.9M	0.845 / 0.800 / 0.835	3.70 / 3.77 / 3.73	0.133 / 0.250 / 0.267
upper (LoRA, layers 24-35 only)	20.0M	0.835	2.63	0.300
persona_only (LoRA on the 320 persona rows alone)	59.9M	0.795	3.75	0.200
math_only (LoRA on the 117 math rows alone)	59.9M	0.815	1.00	0.000
merged (persona_only + math_only LoRAs, cat-concatenated)	119.7M	0.725	3.72	0.150
full (full fine-tune, no adapter, lr 1e-5)	3.09B	0.840	2.42	0.300

Findings: three DoRA seeds set the noise floor (about +-0.02 accuracy, +-0.07 coherence fail; voice +-0.04). Persona needs real capacity spread across the network (IA3 and upper-layers-only lose a third to a half of the voice); full fine-tuning keeps accuracy and loses a third of the voice; DoRA is indistinguishable from LoRA once averaged over seeds; concatenating a persona-only and a math-only LoRA is the one clearly worse arm. Adapter size and type are not the lever.

006: Shakespeare, a persona with a model-free verifiable form

Same recipe as D320 (self-distilled math + persona rows), teacher = Claude Haiku 4.5, persona = sonnets (14 lines, ABAB CDCD EFEF GG, model-free-checked before the judge ever sees them; src/verse_metric.py in the repo).

stage	accuracy	voice (open)	sonnet_ok	rhyme rate	coherence fail
base_prompted (Shakespeare)	0.880	1.02	0.00	0.00	0.00
sonnet sft	0.825	2.10	0.15	0.41	0.60

Accuracy is statistically tied with Gandalf's D320 (p=1, McNemar) -- the format-native-math recipe transfers. Voice and coherence trail Gandalf's; the model nails line count (87%) and syllable count (99%) almost every time but only 41% of rhyme pairs land -- rhyme, not meter, is the bottleneck.

Full write-up: harvard-cs2881f26/hw1-mangini-schwartz-tan, experiments/005_adapter_config/ and experiments/006_shakespeare/.

Usage: PeftModel.from_pretrained(base, "Noaleetz/ai-kitchen-005-006", subfolder="005/dora") (subfolder="005/full" is a full model, not an adapter -- load with AutoModelForCausalLM.from_pretrained instead).