tslm-nov26

OpenTSLM Flamingo model (OpenTSLMFlamingo) built on google/gemma-3-4b-pt, from model_factory training run 02102bec679e4f3d985a884d3f5b2bdf, step 80000. The Gemma backbone was frozen during training; the time-series encoder, perceiver resampler and gated cross-attention layers were trained.

model.safetensors holds the module state_dict (5,241,692,050 parameters, torch.bfloat16, torch.float32). The Flamingo wrapper registers some modules under two names; those 188 alias keys are not stored twice. Their mapping (alias -> stored key) is in the safetensors metadata under aliases:

import json
from safetensors import safe_open
from safetensors.torch import load_file

state_dict = load_file("model.safetensors")
with safe_open("model.safetensors", framework="pt") as f:
    aliases = json.loads(f.metadata()["aliases"])
state_dict.update({alias: state_dict[key] for alias, key in aliases.items()})

License

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. This model is a Model Derivative of Gemma and is distributed under the same terms, including the Gemma Prohibited Use Policy.

Files

• model.safetensors: weights (see above).
• config.json: OpenTSLM Flamingo settings (time-series encoder, perceiver resampler, cross-attention spacing) and the Gemma 3 language model config under language_model_config.
• generation_config.json, tokenizer.json, tokenizer.model, tokenizer_config.json, special_tokens_map.json, added_tokens.json: the Gemma 3 tokenizer with the Flamingo tokens <image>, <image_soft_token> and <|endofchunk|> added.

Loading the model needs the OpenTSLMFlamingo class from the OpenTSLM code.