MaxDGUPTA/uma-cognitive-llm-checkpoints

UMA cognitive-LLM fine-tuning checkpoints

LoRA adapter checkpoints for the cognitive-model-distillation pipeline (UMA -> Qwen3-4B-Base), covering both arithmetic domains and including the per-epoch trajectory of every human fine-tuning stage. Raw inference outputs and figure scripts live in the companion code repo (github.com/MaxDGU/UMA_PR02, branch main_max, see eval/outputs/FIGURE_DATA_MAP.md).

All checkpoints are PEFT LoRA adapters. Directories under */instruct_* load on Qwen/Qwen3-4B-Instruct-2507; everything else loads on Qwen/Qwen3-4B-Base:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

sub = "fractions/distill_humanft/best"  # any directory below
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Base", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "MaxDGUPTA/uma-cognitive-llm-checkpoints",
                                  subfolder=sub).merge_and_unload()
tok = AutoTokenizer.from_pretrained("MaxDGUPTA/uma-cognitive-llm-checkpoints", subfolder=sub)

Each directory carries its train_args.json (full launch config) and history.json (per-epoch train/val NLL). Optimizer states are omitted.

Layout

Path	Contents
fractions/distill_lora	Distillation checkpoint (996-learner UMA panel), init for human FT
fractions/distill_humanft/best	Published distill+humanFT model (1 epoch, best-by-val-NLL)
fractions/panel1000_distill_lora	Panel-scaling arm: distillation on all 1,000 UMA learner profiles, size-matched to the 25-learner panel (468,000 train rows, 3,657 updates, same recipe; best-by-val-loss = final). Profiles 92, 93, 97, 98 were re-simulated with the original code (c=4, matching their saved neighbors)
fractions/panel1000_distill_humanft/seed_42..45	Distill+humanFT for the panel above, 4 human-FT seeds (1 epoch each). Results: github.com/MaxDGU/UMA_PR02 eval/outputs/panel_scaling
fractions/panel1000full_distill_lora	Full-data arm: all 1,000 UMA learner profiles x all 20,720 problems (18.65M train rows, 145,688 updates, 1 epoch, eff. batch 128). Learner tag on 50% of training prompts (--train_prompt_student_mode dropout). Evaluate with SPACED problem prompts (Solve this fraction problem: 3/5 + 1/4=?): this long-trained model is sensitive to the unspaced form
fractions/panel1000full_distill_klhumanft/seed_42..45	KL-anchored human FT (--kl_beta 1.0, reference = the distill checkpoint above) on the same 288 Siegler 2011 rows, 1 epoch, 4 seeds. Results: github.com/MaxDGU/UMA_PR02 eval/outputs/panel_scaling
fractions/distill_humanft_5ep/epoch_1..5, best	5-epoch trajectory of the same config (deterministic retrain; best = epoch 3)
fractions/humanft_frombase/epoch_1,2,3,5, best	Human FT directly from base (no distill; best = epoch 3)
decimals/distill_lora	Decimal distillation checkpoint (BSS-mix), init for human FT
decimals/distill_humanft/best	Published decimal distill+humanFT model (3 epochs, best = epoch 3)
decimals/humanft_frombase/epoch_1,2,3,5, best	Decimal human FT from base (best = epoch 4)
decimals/cv_m1_trainF1/epoch_1..3, best	Problem-disjoint CV fold model M1 (trained on fold 1; best = epoch 3)
decimals/cv_m2_trainF2/epoch_1..3, best	CV fold model M2 (trained on fold 2; best = epoch 2)
fractions/instruct_distill_lora/best	Instruct-2507 fraction distillation checkpoint
fractions/instruct_distill_humanft/epoch_3	Published Instruct-2507 fraction distill+humanFT model (lr 5e-5, epoch 3)
decimals/instruct_distill_lora/best	Instruct-2507 decimal distillation checkpoint
decimals/instruct_distill_humanft/best	Published Instruct-2507 decimal distill+humanFT model (best = epoch 3)

Training data: fractions = Siegler 2011 8-problem human responses (SP2013 16-problem evaluation is problem-disjoint); decimals = BSS2021 92-subject, 12-problem human responses (CV folds are problem-disjoint splits thereof). Seed-variance runs (seeds 43-46) remain on the cluster; ask if needed.