hinglish-relatedness-sbert

A sentence-transformers model for semantic relatedness between romanized Hindi ("Hinglish") words and their English counterparts — fine-tuned from l3cube-pune/indic-sentence-bert-nli with LoRA + mined hard negatives. Built for Warmer, a semantic word-guessing game, where it ranks a player's guess against a secret word by cosine similarity.

Task: given two short romanized-Hindi and/or English words/phrases (1-5 tokens), score how related they are — not just synonymous. "weather" and "forecast" should score high (domain relatedness), matching a WordSim-353-style relatedness notion rather than a strict SimLex-999 similarity notion. This is a relatedness model, not a pure-synonymy model.

Why this exists

Off-the-shelf multilingual sentence embedders handle romanized Hindi (Latin-script, not Devanagari) poorly — a 9-candidate bake-off found every general multilingual model scored near-random on this input form; l3cube-pune/indic-sentence-bert-nli (itself trained on code-mixed Hindi+English text) was the best off-the-shelf option at Spearman 0.639 on a 25-secret production benchmark. Two straightforward attempts to fine-tune it further by adding more training data (543 pairs, then 756/1028 pairs) both made held-out performance worse — the classic small-fine-tune overfitting signature. A follow-up diagnostic swept the training METHOD instead of data volume: 13 capacity-control configs (full fine-tune at several learning rates, frozen transformer layers, a fresh randomly-initialized projection head, LoRA at rank 8/16/32) on the same fixed dataset. Every full-parameter and frozen-layer config stayed within noise of the off-the-shelf baseline; every LoRA config beat it. Layering mined hard negatives onto the best LoRA config (rank 8) pushed held-out performance further still. The bottleneck was optimization capacity, not data.

Evaluation

Field-leading claim, on a public, dedicated benchmark: this model was compared against 7 off-the-shelf alternatives (including 2 released after this project's original bake-off) on a public, scrubbed, 50-secret held-out set — this model leads all 7 on both metrics, CI-significant in all 14 comparisons, zero exceptions (dim1 semantic-ranking Spearman 0.675 [95% CI 0.601, 0.741]; dim2 cross-language pass-rate 0.740 [0.62, 0.86]). Full comparison table and methodology: gauravgandhi2411/hinglish-relatedness-benchmark (a public dataset published alongside this model — the reference romanized-Hinglish relatedness benchmark this field didn't previously have). That benchmark's own dev/test split achieves dev↔test Pearson r = 0.920 (vs. this earlier ~0.12 problem below), so checkpoint-selection decisions built on it are now trustworthy in a way the original held-out split wasn't.

Original held-out generalization claim (frozen train/dev/test split of the game's secret-word schedule, 60 words, 33/11/16 train/dev/test, plus 5 independently-documented hardest-known cases forced into test — kept below for historical continuity with this model's original ship decision; superseded as the field-leading claim by the public benchmark above):

Confidence intervals use a paired bootstrap (the same resampled test items scored by both models in each draw) rather than comparing two independent confidence intervals — this removes shared variance and is the statistically appropriate test for "does the candidate beat the baseline on identical items." The marginal (non-paired) confidence interval on the baseline alone is considerably wider ([0.223, 0.632]) because that ranking metric's independent sample size is only 8 word-clusters — a real limitation of that smaller eval; the public 50-secret benchmark above (42 dim1-eligible secrets) exists specifically to fix this.

Production wrapper (Warmer's actual deployment — generator-side, not a model change)

The numbers above are the RAW model, loaded exactly as shown in Usage below — they are unchanged by anything in this section. Warmer additionally ships two small, build-time-only mitigations in its generator pipeline (not in these model weights) that further improve its own measured relatedness quality on top of this model. Anyone loading this checkpoint directly via sentence-transformers gets the raw numbers above; the numbers below only apply inside Warmer's generator, where both mitigations are applied automatically.

1. Targeted Devanagari-embedding patch (spec-residual-attack.md lever L2): for a small, confirmed list of romanized words whose Latin-script spelling collides with unrelated English subword tokens (e.g. junoon fragments into after+noon), the generator feeds the word's Devanagari spelling to this model's encoder instead of the romanized form — the base tokenizer handles Devanagari script far better. Every other layer (secret matching, hints, on-screen text) still uses the romanized spelling; only the embedding input changes, for a short, fixed list of confirmed collision words (one per line, so as not to imply any relatedness pairing between them):

   • junoon
   • registan
   • jheel
   • udaasi

   This is deliberately narrow, not a general front-end: applying the same substitution to the ~5,200 other Aksharantar-dictionary-coverable vocabulary words was tested at full scale and found to CI-significantly regress dim1 (0.6568 → 0.4731, 19 previously-correct pairs broken) — this model is fine-tuned on romanized text, so broad Devanagari substitution discards that learned representation for words that were never tokenization-pathological in the first place. The narrow, 4-word form is the only version with evidence behind it.

2. Ranking-time commonsense override (lever L5, non-fine-tuning): for a small, curated set of 8 (secret, related-property-word) pairs capturing world-knowledge associations pure co-occurrence embedding structurally under-ranks (e.g. tarbooz/watermelon → lal/red — you know this from experience, not from the two words co-occurring in text), the generator's rank computation is told the fact directly at puzzle-build time: a single array-element splice forcing that one word's rank into the "scorching" band, verified to have zero effect on any other word or secret.

Combined result on the same public 34-secret benchmark: dim1 Spearman 0.6568 → 0.7170 (+0.0602, 95% CI [-0.0017, +0.1435] — directionally positive, not independently CI-significant alone); dim2 cross-lingual pass rate 0.7273 → 0.8485 (+0.1212, 95% CI [+0.0303, +0.2424] — CI-significant). Regression floor: strictly zero broken pairs on both metrics (10 fixed / 0 broken dim1; 4 fixed / 0 broken dim2) — nothing that worked before stopped working. Both levers are zero runtime cost (small dictionary lookups at build time, no extra inference). Full methodology and numbers: generator/evals/finetune/data/probe_l2_patch_l5_combined.eval_result.json and generator/evals/finetune/data/verify_ship_l2_l5_e2e.json (source repo).

Method

• Base: l3cube-pune/indic-sentence-bert-nli (SBERT, mean pooling, 768-dim, BERT-base architecture, CC-BY-4.0).
• LoRA: rank 8, alpha 16, dropout 0.1, applied to the query/value projections in every self-attention layer only (the base weights stay frozen — 0.12% of parameters trainable). This was found to be the decisive fix after full-parameter fine-tuning regressed twice.
• Objective: joint MultipleNegativesRankingLoss (anchor/positive/mined-hard-negative triples) + CosineSimilarityLoss (graded pairs), 4 epochs, learning rate 2e-4, batch size 16, seed 42.
• Hard-negative mining: for each anchor/positive training pair, one additional negative was mined — a vocabulary word the off-the-shelf (pre-fine-tune) model already ranked moderately close to the anchor (rank 50-500 out of ~10k words) — genuinely confusable rather than a random easy negative. This alone outperformed simply doubling the in-batch negative count.

Training data

756 (anchor, positive) contrastive pairs / 1028 graded similarity pairs (0.05-0.95 scored), from 5 sources, all commercial/redistribution-clean:

Two candidate sources considered during dataset research were excluded and contributed nothing to this model: HingCorpus (CC-BY-NC-SA-4.0 — non-commercial, incompatible with a CC-BY release) and IndoWordNet (license text states "provided openly for the purpose of research" — research-only). Every training record was independently checked to exclude any word from the eval's held-out test/dev split, verified by an automated assertion + test before training.

Attribution

This model is a derivative of l3cube-pune/indic-sentence-bert-nli (L3Cube-Pune, CC-BY-4.0). Training data derives in part from IndicCorp v2 and Aksharantar (both AI4Bharat). LLM-augmented training pairs and one transliteration-fallback path were generated locally using Qwen3-8B (Alibaba, Apache-2.0). Released under CC-BY-4.0 — attribution to the above sources required for reuse/redistribution.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("gauravgandhi2411/hinglish-relatedness-sbert")

words = ["namaste", "cricket", "aeroplane"]
embeddings = model.encode(words)

similarities = model.similarity(embeddings, embeddings)
print(similarities)

Intended use

Ranking a small candidate vocabulary (thousands to tens of thousands of short romanized-Hindi/English words) against a query word by relatedness — e.g. a word-guessing game's "how warm is this guess" signal. Not evaluated for long-form sentence similarity, retrieval-augmented generation, or Devanagari-script input (romanized/Latin-script Hindi only, matching the base model's training distribution).

Limitations (reported honestly, not smoothed over)

• Small held-out eval set. The ranking-correlation metric above is computed over only 8 independent word-clusters (73 individual scenarios) — the marginal 95% confidence interval on the baseline alone is [0.223, 0.632], i.e. wide. The paired-bootstrap comparison above is the statistically correct way to read "does this model beat the baseline," but readers should not treat 0.7044 as a precise point estimate on a larger benchmark.
• The training loop's own checkpoint-selection metric barely predicts held-out performance on the ORIGINAL small split — across the diagnostic sweep that produced this model, the dev-set metric used for early stopping correlated with the held-out test metric at only Spearman 0.12 there. Fixed on the larger public benchmark (dev↔test Pearson r = 0.920, see Evaluation) — a future re-training run should select checkpoints against that split, not the original one.
• The hardest-known cases are improved but not solved. 5 independently-documented hardest word pairs went from a Spearman correlation of ~0 to 0.562 on the ranking metric — a large improvement, but still below the 0.7044 full-test-set average, i.e. these remain the relatively harder cases even after fine-tuning.
• Small training set (1741 total pairs after hard-negative augmentation) relative to a ~110M-parameter base model. A learning-curve check found held-out performance still rising at 100% of available data — more (real, in-domain) training data would likely help further; this was not a data-quality ceiling, LoRA's low effective capacity is what made the existing data usable without overfitting.
• A residual weakness on a specific, characterized class of short/rare romanized words — the public benchmark localizes this to the emotion/nature domains specifically (junoon/passion, registan/desert being the clearest cases). Root cause: these words' own embeddings are dominated by unrelated English words sharing a coincidental subword-tokenization fragment (e.g. junoon→afternoon/noon/june) — a base-tokenizer artifact, distinct from a separate associative/world-knowledge gap (e.g. imli/tamarind→sour) this distributional approach also structurally under-ranks. A full multi-lever research pass (spec-residual-attack.md) attacked both: PROVEN DEAD (measured, do not re-attempt without new evidence) — higher LoRA capacity (overfits, degrades monotonically with capacity) and an alternative base model with different romanized-Hindi tokenization (MuRIL shares a byte-identical vocab with the current base, mathematically incapable of changing anything). Also proven dead: a general Devanagari-embedding front-end applied broadly rather than targeted (CI-significant dim1 regression at scale, see the Production wrapper section above). Shipped as a build-time mitigation (Production wrapper section above): a narrow, 4-word targeted Devanagari-embedding patch for the tokenization-collision words, plus a ranking-time knowledge override for 8 curated commonsense-association words — CI-significant dim2 improvement, zero net regression, no model retrain. Still open for a future session: knowledge distillation from a larger instruction-tuned teacher (bounded, zero-regression gain measured on trained words only, not yet tested at scale) and commonsense-shaped training data (held-out generalization test was inconclusive — 5/8 improved, 3/8 regressed, n=8 too small to resolve). This is a triaged, characterized boundary with a documented path back in, not a closed or believed-unfixable limitation. Full detail: this project's docs/known-limitations.md entry 20 and spec-residual-attack.md.

Citation

If you use this model, please cite the base model and the data sources above per their licenses. This fine-tune itself has no associated paper.