Llama-3.2-3B-Instruct-SFT-ScienceWorld
PRIVATE — internal research checkpoint. Do not redistribute.
| Task | ScienceWorld |
| Stage | SFT (behavior-cloning) initialization |
| Base | meta-llama/Llama-3.2-3B-Instruct (Built with Llama) |
| Init for this run | models/llama-3.2-3b-instruct |
| Source checkpoint | rl/ckpts/sw_sft_llama3.2-3b/global_step_198 (global_step 198) |
| Format | bf16 safetensors, merged from hf-dir -> bf16 cast |
| Params | 3,212,749,824 |
| Weight drift vs. init (mean rel-L2) | 4.69e-03 (max 9.40e-03, 226 / 255 tensors changed) |
| Reload bit-exact check | True |
Optimizer state is not included (inference/eval only; the fp32 FSDP shards
stay on HiPerGator /blue). Load with AutoModelForCausalLM.from_pretrained(..., torch_dtype=torch.bfloat16)
or vLLM; the chat template is the stock Llama-3 template shipped in chat_template.jinja.
Licensed under the Llama 3.2 Community License; derivative of Meta Llama.