astrobridge-model-v8

AstroBridge Captioner

n-modality (image + spectra) astronomy captioner. LoRA adapter + fusion stack trained on top of a frozen Qwen/Qwen3.5-9B. The base model itself is NOT included here — load it fresh from Qwen/Qwen3.5-9B and apply this adapter on top.

How to load

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.5-9B", dtype=torch.bfloat16, trust_remote_code=True
)
llm = PeftModel.from_pretrained(base, "UniverseTBD/astrobridge-model-v8")
tokenizer = AutoTokenizer.from_pretrained("UniverseTBD/astrobridge-model-v8")

# middle.pt (fusion stack: projectors/modality_identity/qformer/adapter) needs the captioner
# package's FusionStack class to reload — see captioner/model/captioner.py and
# captioner/train/stage1.py's run_stage1 for how it's constructed and wired to the LLM.

Training info

• config_hash: 33ed49969a8fce7d
• quantization: None
• git_sha: 392fcf007fd11d12d04ffd13868d0c19dc25d190
• tier_histogram: {"single": 5564}

Eval (groundedness gate)

{
  "per_modality": {
    "image": {
      "shuffle_test": {
        "modality": "image",
        "n": 39,
        "mean_edit_distance": 246.84615384615384,
        "null_result": false
      },
      "ablation_test": {
        "modality": "image",
        "n": 84,
        "fraction_caption_changed": 1.0,
        "null_result": false
      },
      "diversity_test": {
        "modality": "image",
        "n": 120,
        "distinct_fraction": 1.0,
        "top1_share": 0.008333333333333333,
        "most_common_caption": "This object appears as a compact, smooth, and slightly elongated source oriented diagonally from the lower left toward the upper right. It features a bright, concentrated yellowish-white center that f",
        "null_result": false
      }
    },
    "spectra": {
      "shuffle_test": {
        "modality": "spectra",
        "n": 26,
        "mean_edit_distance": 181.15384615384616,
        "null_result": false
      },
      "ablation_test": {
        "modality": "spectra",
        "n": 72,
        "fraction_caption_changed": 1.0,
        "null_result": false
      },
      "diversity_test": {
        "modality": "spectra",
        "n": 120,
        "distinct_fraction": 1.0,
        "top1_share": 0.008333333333333333,
        "most_common_caption": "This optical spectrum of a galaxy in the local universe displays a red continuum characteristic of an early-type host with a prominent 4000 \u00c5 break. Superimposed on the continuum are narrow nebular em",
        "null_result": false
      }
    },
    "lightcurve": {
      "shuffle_test": {
        "modality": "lightcurve",
        "n": 9,
        "mean_edit_distance": 46.22222222222222,
        "null_result": false
      },
      "ablation_test": {
        "modality": "lightcurve",
        "n": 44,
        "fraction_caption_changed": 1.0,
        "null_result": false
      },
      "diversity_test": {
        "modality": "lightcurve",
        "n": 91,
        "distinct_fraction": 1.0,
        "top1_share": 0.01098901098901099,
        "most_common_caption": "The optical light curve spans approximately 170 days in the $g$ and $r$ bands, with the brightest sampled detections occurring near day 10.0 ($g = 17.19 \\pm 0.07$ mag, $r = 17.12 \\",
        "null_result": false
      }
    }
  }
}