base_model string | system_prompt string | prompt_source string | gen_max_new_tokens int64 | keep_min_dev float64 | n_pairs int64 | created_utc timestamp[s] | pairs list | eval_prompts list | base_eval_out list |
|---|---|---|---|---|---|---|---|---|---|
google/gemma-4-E2B-it | You are a helpful AI assistant that answers in Nepali. | OpenAssistant/oasst1 (apache-2.0), English conversation openers | 128 | 0.7 | 468 | 2026-09-18T19:47:55 | [["What are the rules for writing a Haiku? Can you write me an example?","हाइकू लेख(...TRUNCATED) | ["How do inductors work?","Let’s imagine this: You are a linux shell using sh. My first command is(...TRUNCATED) | ["Inductors are fundamental passive electronic components that store energy in a **magnetic field** (...TRUNCATED) |
Nepali SFT pairs for Gemma 4 E2B
468 (English prompt -> Nepali answer) pairs, the exact training data behind litert-community/gemma-4-E2B-it-nepali-lora.
Published so the training notebook can skip a ~13 minute generation step and so anyone reproducing it evaluates on the same held-out split.
Provenance
- Prompts: English conversation openers from
OpenAssistant/oasst1(Apache-2.0, human-written), filtered torole == "prompter",parent_id is None,lang == "en". - Targets: generated by
google/gemma-4-E2B-ititself, prompted with "You are a helpful AI assistant that answers in Nepali.", then quality-gated at Devanagari ratio >= 0.7.
The targets are model output, not human translation. Any factual error or awkward phrasing the base model produces is present in the data. This dataset is for reproducing a language-switching behaviour experiment, not a reference corpus for Nepali.
Measured properties
| training pairs | 468 |
| Devanagari ratio of targets | mean 0.859, min 0.700 |
| held-out eval prompts | 60 |
| base-model Devanagari on those | 0.000 |
| train/eval prompt overlap | 0 |
The 0.000 baseline is the point: without a system prompt the base model answers these English prompts entirely in English, so any Devanagari after training is attributable to the adapter.
Why the gate is 0.70 and not 0.95
OASST1 contains many programming questions, and a correct Nepali answer to "how do I use
tf.where?" keeps its identifiers in Latin script. Since the metric counts every alphabetic
character, code and proper nouns legitimately drag good answers down to ~0.7. At a 0.90 threshold
only 26% of generations survived; at 0.70, 78% do.
Schema
A single JSON object in nepali_pairs.json:
| key | type | meaning |
|---|---|---|
base_model |
str | model that generated the targets; asserted on load |
system_prompt |
str | prompt used to elicit them; asserted on load |
pairs |
list[[str, str]] | [english_prompt, nepali_target] |
eval_prompts |
list[str] | held-out prompts, never trained on |
base_eval_out |
list[str] | base-model answers to those, for the before/after baseline |
keep_min_dev |
float | quality-gate threshold applied |
gen_max_new_tokens |
int | generation length cap |
created_utc |
str | when this file was packaged for publication |
import json
from huggingface_hub import hf_hub_download
d = json.load(open(hf_hub_download(
repo_id="litert-community/gemma4-e2b-nepali-sft-pairs", filename="nepali_pairs.json", repo_type="dataset")))
pairs = [tuple(x) for x in d["pairs"]]
The companion training notebook consumes this automatically with USE_CACHED_PAIRS = True.
License
Apache-2.0, matching the OASST1 prompts and the Gemma 4 outputs.
Notice
This repository contains an educational demonstration artifact created for the LiteRT-LM fine-tuning Colab tutorial to illustrate end-to-end LoRA fine-tuning and .litertlm conversion. It is provided 'AS IS' without warranty of any kind and is not an official general-purpose benchmark or production model release.
- Downloads last month
- 82