sudarshanregmi/timethink-base

TimeThink (Base)

Untrained starting checkpoint for TimeThink: a fresh Qwen3-8B backbone plus a time-series encoder ported from ChatTS-8B — only the ts_encoder weights are taken from ChatTS-8B; the language-model backbone is the base Qwen3-8B (not ChatTS's fine-tuned backbone). The model has not been fine-tuned; this is the checkpoint that scripts/sft.sh trains from.

Trained models: timethink-sft · timethink-rl

Load

from transformers import AutoModelForCausalLM, AutoProcessor

model = AutoModelForCausalLM.from_pretrained(
    "sudarshanregmi/timethink-base", trust_remote_code=True)
processor = AutoProcessor.from_pretrained(
    "sudarshanregmi/timethink-base", trust_remote_code=True)

Requires transformers>=4.55,<5.

Evaluation data: sudarshanregmi/timethink