FlameLedger-Encoder
semantic search for NASA fire safety reports
2 min readupdated
A text encoder by Team Qubit (built for the NASA Space Apps Challenge 2026) based on google/embeddinggemma-2 (740M total parameters, 271M in the text backbone). The model encodes English queries and scientific texts into vectors (embeddings) for semantic search across NASA technical reports on microgravity combustion and spacecraft fire safety. Unlike the multimodal base model, only the text tower is kept here, fine-tuned on synthetic questions generated from the NTRS database. Useful for developers building RAG (retrieval-augmented generation) pipelines and research assistants for aerospace documentation.
Key features
- Matryoshka embeddings support: 768-dimensional vectors can be truncated to 512, 256, or 128 dimensions without retraining to save memory.
- Only the text backbone (271M parameters) was fine-tuned in FP32 precision — the vision and audio modules from the base model were stripped out.
- Mandatory prefixes: queries and documents must follow the exact
task: search result | query: ...andtitle: ... | text: ...formats, otherwise retrieval quality drops. - The authors recommend pairing the model with BM25 lexical search via RRF (Reciprocal Rank Fusion) rather than running it standalone.
What you can do
- Plug the encoder into the first retrieval stage of a RAG pipeline to navigate English-language NASA and NIST papers on material combustion.
- Build a hybrid search engine (dense embeddings + BM25) to retrieve reports by both semantic meaning and exact alloy codes or test parameters.
- Shrink your vector index down to 256 dimensions with Matryoshka truncation and L2 normalization, keeping nearly the same search accuracy with a smaller storage footprint.
Example
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("rarfileexe/FlameLedger-Encoder")
model.max_seq_length = 512
q = model.encode(["task: search result | query: flame spread in low-speed flow"], normalize_embeddings=True)
d = model.encode(["title: Combustion report | text: Flame spread over thin solid fuel..."], normalize_embeddings=True)
print(q @ d.T)Read more
DetailsOn a synthetic benchmark (150 queries across 39 held-out papers), the Fine-tuned + BM25 hybrid reached Recall@1 0.813 and MRR@10 0.891, compared to 0.787 and 0.873 for the base model with BM25.
Trained on 1,025 triplets extracted from 486 chunks (mostly abstracts) on 1× NVIDIA L4 (24 GB) with a 512-token max sequence length.
Limitations: English only, narrow domain, no validation on human expert queries; running the base model in FP16 is not recommended (use FP32 or BF16).










Comments