EchoSonar

sft/

SFT checkpoint (full fine-tune, step 1503) for the EchoVLM report-generation model.

• sft/llm/ — full HF AutoModelForCausalLM-compatible checkpoint (Qwen3-8B backbone, full fine-tune) with tokenizer included. Load directly with:

  from transformers import AutoModelForCausalLM, AutoTokenizer
  
  model = AutoModelForCausalLM.from_pretrained("daryataratynova8/echosonar", subfolder="sft/llm")
  tokenizer = AutoTokenizer.from_pretrained("daryataratynova8/echosonar", subfolder="sft/llm")
  

• sft/clip_projector.pt, sft/detr_projector.pt — state_dicts for the CLIP and DETR vision-feature projectors (each LayerNorm -> Linear -> GELU -> Linear, mapping vision feature dim to the LLM hidden size). These are not loadable via from_pretrained; load them manually into matching nn.Sequential modules, e.g.:

  import torch
  
  clip_projector_sd = torch.load("clip_projector.pt", map_location="cpu")
  detr_projector_sd = torch.load("detr_projector.pt", map_location="cpu")
  

grpo/

GRPO-tuned checkpoint (grpo_final_exp_d_final_correct), built on top of sft/. GRPO trains a LoRA adapter (r=64) plus the two projectors on top of the frozen SFT LLM — it does not replace the base LLM weights.

• grpo/lora_adapter/ — PEFT LoRA adapter. Merge onto the SFT LLM:

  from transformers import AutoModelForCausalLM
  from peft import PeftModel
  
  base = AutoModelForCausalLM.from_pretrained("daryataratynova8/echosonar", subfolder="sft/llm")
  model = PeftModel.from_pretrained(base, "daryataratynova8/echosonar", subfolder="grpo/lora_adapter")
  model = model.merge_and_unload()
  

• grpo/clip_projector.pt, grpo/detr_projector.pt — projector state_dicts (same shapes/format as sft/, further fine-tuned during GRPO), loaded the same manual way.

• Tokenizer files are identical to sft/llm/, included here for convenience.