SmolLM3-3B-SFT-ScienceWorld
| Task | ScienceWorld |
| Stage | SFT (behavior-cloning) initialization |
| Base | HuggingFaceTB/SmolLM3-3B |
| Init for this run | models/smollm3-3b-agent |
| Source checkpoint | rl/ckpts/sw_sft_smollm3-3b/global_step_198 (global_step 198) |
| Format | bf16 safetensors, merged from hf-dir fp32 -> bf16 cast (per-tensor, login node) |
| Params | 3,337,766,912 |
| Weight drift vs. init (mean rel-L2) | 4.72e-03 (max 1.07e-02, 281 / 326 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.
Use the bundled
chat_template.jinja— it is NOT the stock SmolLM3 template. This checkpoint was behaviour-cloned with a task-specific template that (a) defaults to/no_think, (b) does not pre-fill an empty<think></think>block into the generation prompt, and (c) replaces the long default reasoning instructions with a compact format directive. The model is expected to emit<think>…</think><action>…</action>itself. Loading the stock template instead silently changes the prompt distribution.
Apache-2.0, same as the base model.