| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3.5-27B |
| library_name: transformers |
| pipeline_tag: image-text-to-text |
| tags: |
| - weather |
| - meteorology |
| - radar |
| - sounding |
| - satellite |
| - vision-language-model |
| - image-text-to-text |
| - qwen |
| - isobar |
| --- |
| |
| # Isobar-1 |
|
|
| `Isobar-1` is a weather-specialized vision-language model for radar, sounding, satellite, and forecast-map interpretation. |
|
|
| This release is the merged full model version of a two-stage fine-tune built on top of `Qwen/Qwen3.5-27B`. |
|
|
| ## What This Model Is |
|
|
| - Base model: `Qwen/Qwen3.5-27B` |
| - Release family: `Isobar` |
| - Release: `Isobar-1` |
| - Model type: merged multimodal causal LM / image-text model |
| - Intended use: weather image interpretation, technical meteorology QA, operational weather analysis assistance |
|
|
| ## Training Lineage |
|
|
| `Isobar-1` was produced by merging a stage-2 adapter into the base model. |
|
|
| Training stages: |
|
|
| 1. Broad weather VLM adaptation on `deepguess/weather-vlm` |
| 2. Curated refinement on `deepguess/weather-analysis-sft` |
|
|
| The stage-2 refinement was trained from the stage-1 weather checkpoint rather than from the raw base model, so this release reflects both stages together. |
|
|
| ## Intended Use |
|
|
| Good use cases: |
|
|
| - Interpreting radar imagery |
| - Interpreting soundings and hodographs |
| - Explaining severe-weather setups |
| - Reading forecast maps and model visualizations |
| - Technical weather question answering grounded in images |
|
|
| Not the target for this release: |
|
|
| - General-purpose agent/tool calling |
| - Fully autonomous forecast operations |
| - Non-weather multimodal tasks |
|
|
| A later `Isobar-1-Agent` style release is the better place for dedicated tool-use behavior. |
|
|
| ## Limitations |
|
|
| - This model can still be wrong on edge cases, ambiguous imagery, and unusual regional setups. |
| - It is tuned for weather analysis, not guaranteed forecast verification skill. |
| - It should not be treated as a standalone warning or life-safety authority. |
| - Image quality, missing context, bad timestamps, and incomplete annotation can still degrade output quality. |
|
|
| ## Loading |
|
|
| ```python |
| from transformers import AutoModelForImageTextToText, AutoProcessor |
| |
| model_id = "deepguess/Isobar-1" |
| |
| processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True) |
| model = AutoModelForImageTextToText.from_pretrained( |
| model_id, |
| trust_remote_code=True, |
| torch_dtype="auto", |
| device_map="auto", |
| ) |
| ``` |
|
|
| ## Attribution |
|
|
| This model is based on `Qwen/Qwen3.5-27B`. |
|
|
| The release name `Isobar-1` is a downstream branding name for this fine-tuned derivative. It is not an official Qwen release. |
|
|
| ## License |
|
|
| This release is provided under `Apache-2.0`, consistent with the upstream base model license. See the upstream model card and license terms for the base model as well. |
|
|
