Quanfire Government Embedding — embed-gov-indic (cross-lingual, Indian government press releases)

A cross-lingual retrieval adapter for Indian government press-release text across 16 Indian languages. It is a LoRA adaptation over a frozen intfloat/multilingual-e5-small (MIT) base — a 2.4 MB adapter, 384-dimensional normalized vectors, max_length 256 — trained on Press Information Bureau (PIB) press releases, which are published as the same release in many languages and so provide naturally-parallel cross-lingual signal.

The point of this model is cross-lingual retrieval: a query in one Indian language finding the passage about the same government release in another language. It is a specialist and a showcase of clean-provenance multilingual retrieval, not a general model — and this card shows you exactly what it does and does not do, measured.

• Framework & code: github.com/quanfire-ai/quanfire-multilingual-embedding (Apache-2.0)
• Weights licence: Apache-2.0. The adapter is ours; the training text is PIB press-release content reused under PIB's reproduction policy (free reproduction, attribution, no NonCommercial, no ShareAlike) — attribution is given below, and the model is non-reconstructive (it emits vectors, never the source text).
• Internal run: gov-indic-e3 · base e5-small · rank 32 / alpha 64, LoRA on query,value · 589,824 adapter params · 1 epoch, lr 1e-4, batch 256 (bf16, CUDA)

What it is for

Cross-lingual and in-language retrieval over Indian government press releases — search a corpus of releases with a query in any of the supported languages, and rank passages regardless of which language they are written in. It embeds a query and a passage into the same 384-d space; cosine similarity ranks.

Scope — read this before you use it

	Validated?
Cross-lingual retrieval over government press-release text (16 Indian languages)	✅ Yes — significant, measured below
Other domains (legal, finance, news, conversational, product)	❌ Not validated — this is a government-press-release specialist
The lowest-resource languages here (Khasi, Nepali, Manipuri)	⚠️ Thin — very few eval examples; treat their numbers as indicative only

The base multilingual-e5-small is already a capable multilingual retriever, so the honest framing is: this adapter adds a significant, broad cross-lingual gain on government-domain text on top of an already-multilingual base. Absolute Recall@1 is modest (retrieving the one right passage out of ~1,800 candidates across 16 languages is hard); the right passage lands in the top-10 about 74% of the time.

Results (held-out, scored on CUDA)

In-distribution — 1,800 held-out cross-lingual pairs (anchor and positive in different Indian languages, same release). The published base is the only honest baseline; the adapter is scored on the same held-out pairs.

Metric	base e5-small	embed-gov-indic	change
Recall@1	0.1836	0.2349	+27.9%
Recall@5	0.516	0.636	+23%
Recall@10	0.649	0.744	+15%
MRR	0.336	0.408	+21%

The Recall@1 improvement is statistically significant — the 95% confidence intervals are disjoint (base [0.165, 0.204] → adapter [0.214, 0.257]).

Per positive-language (Recall@1, before → after), largest eval groups:

Lang	before	after	Lang	before	after
English	0.167	0.297	Malayalam	0.175	0.223
Urdu	0.206	0.313	Punjabi	0.175	0.204
Hindi	0.324	0.405	Bengali	0.179	0.221
Tamil	0.150	0.240	Telugu	0.214	0.235
Kannada	0.142	0.248	Odia	0.138	0.160
Marathi	0.219	0.263	Assamese	0.065	0.083

Gains are broad across languages; a few very-low-count languages (Gujarati −0.02, Manipuri, Khasi) move by noise-level amounts and should not be over-read.

Licence & provenance

• Base: intfloat/multilingual-e5-small (MIT).
• Training data: Press Information Bureau (PIB, pib.gov.in) press releases, reused under PIB's stated reproduction policy — free reproduction with source acknowledgement, no NonCommercial and no ShareAlike clause. Embedded third-party material is excluded. Attribution: Source — Press Information Bureau (pib.gov.in), Government of India.
• Weights: Apache-2.0. Because the model is non-reconstructive (it emits 384-d vectors and never reproduces the source text), and the source is reproduction-permitted with attribution (satisfied here), no ShareAlike floor applies to the weights.

How it was built (honest, in three runs)

• e1 — mixed pair kinds, language-imbalanced: +8.7%, not significant, with regressions in low-resource languages.
• e2 — cross-lingual title↔body pairs only, language-balanced: +16.2%, regressions fixed, but still not significant on a small eval.
• e3 (this release) — a larger balanced corpus and a larger held-out eval: +27.9%, statistically significant.

The corpus was built by re-sourcing PIB releases directly (never via third-party scraped compilations), joining each release to its sibling-language versions, and mining cross-lingual title↔body pairs with a lexical-overlap cap so a gain reflects meaning, not string matching.

Intended use & limits

Use it to retrieve/rank Indian government press-release passages, including across languages. Do not expect it to transfer to other domains, and treat the lowest-resource languages as indicative. It is a specialist showcase of what clean, provenance-transparent data yields for Indian-language cross-lingual retrieval.