ttthyme-ckpts

Thyme SFT checkpoints from reproducing Table 5 "+Only Last Round" of Thyme paper.

See full repro notes at Irisicy4/Thyme-projects @ refactor-reproduce-notes.

Checkpoints

Folder	Config	Train epochs	train_loss	Notes
SFT-baseline-bs128-seed42/checkpoint-4251	effective batch 128 (matches paper)	3	0.274	main result
SFT-baseline-bs64-DEPRECATED-seed42/checkpoint-8499	effective batch 64 (half-paper, setup error)	3	0.206	partially eval'd
SFT-baseline-bs64-partial-seed42/checkpoint-1500	bs64 — partial	0.5	n/a	killed mid-training
SFT-no123-bs64-partial-seed42/checkpoint-2000	bs64 — partial — no123 ablation	0.7	n/a	killed mid-training

Base model

All checkpoints are fine-tunes of Qwen/Qwen2.5-VL-7B-Instruct. Load with:

from transformers import AutoModelForCausalLM, AutoProcessor
mp = "Icey444/ttthyme-ckpts/SFT-baseline-bs128-seed42/checkpoint-4251"
model = AutoModelForCausalLM.from_pretrained(mp, trust_remote_code=True)
processor = AutoProcessor.from_pretrained(mp)

Or with VLMEvalKit + the Thyme code-execution sandbox loop (see repro notes §13 in the Irisicy4 repo branch above).