Instructions to use sraivante/TARA-English-Tutor-3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use sraivante/TARA-English-Tutor-3B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="sraivante/TARA-English-Tutor-3B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("sraivante/TARA-English-Tutor-3B")
model = AutoModelForCausalLM.from_pretrained("sraivante/TARA-English-Tutor-3B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use sraivante/TARA-English-Tutor-3B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sraivante/TARA-English-Tutor-3B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sraivante/TARA-English-Tutor-3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/sraivante/TARA-English-Tutor-3B

SGLang

How to use sraivante/TARA-English-Tutor-3B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "sraivante/TARA-English-Tutor-3B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sraivante/TARA-English-Tutor-3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "sraivante/TARA-English-Tutor-3B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sraivante/TARA-English-Tutor-3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use sraivante/TARA-English-Tutor-3B with Docker Model Runner:
```
docker model run hf.co/sraivante/TARA-English-Tutor-3B
```

TARA-English-Tutor-3B / training_data /README.md

sraivante

Publish TARA English Tutor 3B model with training data

24e1610 verified 15 days ago

preview code

Raw History Blame Contribute Delete

6.58 kB

metadata

pretty_name: TARA English Tutor Instructions
language:
  - en
license: apache-2.0
task_categories:
  - text-generation
size_categories:
  - 1K<n<10K
tags:
  - education
  - english-tutor
  - instruction-tuning
  - synthetic
  - structured-output
  - json
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/validation.jsonl

TARA English Tutor Instructions

An English tutor dataset for Nursery, LKG, UKG and Classes 1–5, prepared from original curriculum templates and help-seeking examples. It teaches short grade-conditioned answers, gentle English correction, speaking practice, explained vocabulary and trusted-adult responses.

It accompanies TARA English Tutor 3B, a SmolLM3 QLoRA fine-tune. Every record retains teacher_reviewed: false. This experimental release does not claim educator approval, child readiness or CBSE/NCERT certification.

Split	Records	Curriculum	Help-seeking/safety	Concept groups
Train	1,138	1,010	128	65
Validation	126	110	16	7
Total	1,264	1,120	144	72

Grade	Train	Validation
Nursery	96	22
LKG	156	22
UKG	156	12
Class 1	156	22
Class 2	146	22
Class 3	146	12
Class 4	136	12
Class 5	146	2

Validation is small and uneven across grades, particularly Class 5. Both splits share templates, style and vocabulary policies, so split isolation alone does not establish broad generalization.

Loading and schema

from datasets import load_dataset

data = load_dataset("sraivante/TARA-English-Tutor-Instructions")
print(data["train"].num_rows)       # 1138
print(data["validation"].num_rows)  # 126
messages = data["train"][0]["messages"]

Each row contains:

Field	Meaning
`id` / `group_id`	Stable example identifier and grouping used for split isolation
`grade`	One of the eight level identifiers
`topic` / `skill` / `kind`	Curriculum area, target skill and curriculum or safety record type
`cbse_outcomes`	Project-assigned curriculum labels; not official approval
`messages`	System, user and assistant turns
`response`	Parsed tutor response
`quality`	Original-content declaration, vocabulary warnings and teacher-review status

The assistant response is a JSON string within messages and a parsed object in response. It has exactly five fields:

{
  "answer": "A short direct answer.",
  "practice_question": "One related speaking question?",
  "practice_answer": "A short model answer.",
  "new_words": [{"word": "term", "meaning": "An easy explanation."}],
  "encouragement": "Good thinking!"
}

This illustrates the schema; the data contains concrete questions and answers. The grade profiles permit main answers from 18 words in Nursery to 100 words in Class 5, with one to three explained new words depending on grade. See grade_profiles.json. These are training targets, not guarantees of generated behavior.

Use the model's chat template. TARA's workflow disables extended thinking, masks system/user tokens and trains on assistant tokens only. Computing loss over every token would differ from that workflow. Reproducibility code is included under source/ in the model repository.

Origin and model association

The source is the publisher's CUSTOM_LLM4 project. Its 54 original curriculum concepts are expanded through question variants, grade prompts and encouraging responses. Fixed help-seeking examples add privacy, danger and trusted-adult response patterns. The source declares original content, with no copied textbook passages or collected child conversations.

The deterministic split hashes seed 42 and group_id. All variants of a concept stay in one split. The requested validation fraction is 0.1; actual counts result from assigning whole groups.

Both files reproduce byte for byte from the original cbse_english_tutor_colab.zip preparation bundle and the current local generation source. The notebook rebuilds this dataset before training. However, the actual Colab run's dataset hashes and training manifest were unavailable, so exact historical identity with that run cannot independently be proven. provenance.json records bundle, source-file and data hashes.

The separate TARA 300M scratch model's pretraining corpora are outside this 3B fine-tune release.

Validation and limitations

The release audit found zero schema errors, vocabulary-warning rows, duplicate message records and duplicate IDs. Between splits, it found zero shared concept groups, user prompts or identical messages. Selected email, credential-like-string and phone-number patterns had zero matches; this was not a comprehensive personal-data audit.

All 1,264 records retain teacher_reviewed: false. Automated checks do not establish factual correctness, age appropriateness, diversity or real-world safety. Some synthetic help-seeking prompts concern danger, self-harm or abuse; they teach trusted-adult response patterns.

This small template dataset is for instruction tuning an English-capable model, not language-model pretraining from scratch. It is intended for supervised educational experiments, not ranking children, diagnosing conditions or replacing teachers and responsible adults.

Run python validate_dataset.py after downloading to verify hashes, counts, chat structure and group isolation. dataset_audit.json contains the audit results and scope.

Copyright, licensing and citation

The original project's MIT attribution is retained in SOURCE_LICENSE_MIT. Attribution does not claim exclusive rights over facts, curriculum ideas or independently owned material. SHA256SUMS covers the release files.

@dataset{sraivante2026tara,
  author = {sraivante},
  title = {TARA English Tutor Instructions},
  year = {2026},
  url = {https://huggingface.co/datasets/sraivante/TARA-English-Tutor-Instructions}
}

sraivante
/

TARA-English-Tutor-3B

TARA English Tutor Instructions

Contents

Loading and schema

Origin and model association

Validation and limitations

Copyright, licensing and citation