LiveGraph Gemma 4 31B SFT — linear trace, seed 11
This is a full-parameter BF16 fine-tune of google/gemma-4-31B for the LiveGraph synthetic Boolean-propagation task. The training rendering uses a linearized execution trace. This checkpoint is one seed of a matched representation study; it is not an instruction-tuned general-purpose assistant.
Model details
- Architecture:
Gemma4ForConditionalGeneration
- Checkpoint index metadata: 32,682,372,656 parameters and 62,546,177,752 serialized bytes in two BF16 safetensors shards
- Parameter accounting: the separate GRPO runtime manifest reported 31,395,515,952 parameters after adding 122,429,440 LoRA parameters. The full-checkpoint index retains the upstream base index's 32,682,372,656 logical-parameter convention. Runtime unique-module counts and index metadata can differ because of weight tying and language, embedding, and vision components; the serialized payload and weight map were verified directly.
- Training: 480 unique synthetic items, balanced TRUE/FALSE labels, 8 epochs
- Objective: completion-only supervised fine-tuning
- Context limit used in training: 4,096 tokens; no training row was truncated
- Learning rate / global batch size:
1e-5 / 32
- Seed / final training loss: 11 / 0.703738
- Base config SHA-256:
e967dd38bc5cfd38bd09a995a7bf4a754075df2b46aba68f7fbb5a791e6d8dd1
- Training-data SHA-256:
d5bdea5629c84bbd7fec51f5969e06612b27b817bc302a15045fe7a46e2eb792
Evaluation
Strict exact-label accuracy was measured on a frozen 320-item set spanning depths 1–32, under native thinking, max_tokens=8192, and vLLM FP8 serving. Unparseable completions count as incorrect.
| Evaluation rendering | Strict accuracy | Parse rate |
|---|
| Typed graph | 49.38% | 100.00% |
| Linear trace | 47.19% | 100.00% |
These are protocol-conditional research measurements, not evidence of broad capability improvement. The study used only three seeds and a fixed execution order, so causal claims about representation are not supported.
Use
1from transformers import AutoModelForMultimodalLM, AutoProcessor
2
3processor = AutoProcessor.from_pretrained(".")
4model = AutoModelForMultimodalLM.from_pretrained(
5 ".", dtype="auto", device_map="auto"
6)
Use the supplied chat template and the LiveGraph task prompt format. The model was evaluated only for short TRUE/FALSE completions on synthetic graph instances.
Limitations and risks
The data are synthetic and English-only. Accuracy is near chance on this evaluation, and results may change with decoding, quantization, prompt formatting, or graph distribution. The base model's multimodal and safety properties were not re-evaluated after fine-tuning. Do not use this checkpoint for consequential decisions.
License and provenance
The checkpoint is a modified derivative of Google DeepMind's google/gemma-4-31B, whose local model card declares Apache License 2.0. The full license is included in LICENSE; modification and attribution information is in NOTICE. File hashes are listed in SHA256SUMS.
