ccisd-teks-educator-mistral7b

CCISD TEKS Educator (Mistral-7B LoRA)

A LoRA adapter over mistralai/Mistral-7B-Instruct-v0.2, fine-tuned for instructional-planning language grounded in TEKS, the Texas Essential Knowledge and Skills curriculum standards.

This repository contains only the adapter. It is not a standalone model and will not load without the base model.

Status: experimental and unevaluated. No benchmark results exist for this adapter, and the training run was not recorded at upload time. Treat it as a preliminary artifact rather than something to build on. See Limitations before any use.

What is actually in this repo

Read directly from adapter_config.json and the uploaded weights:

The parameter count is derived from the rank and the Mistral-7B layer dimensions, and it agrees with the weight file: 83,886,080 x 4 bytes = 335,544,320, against an actual 335,604,696 including the safetensors header.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "mistralai/Mistral-7B-Instruct-v0.2"
ADAPTER = "robworks-software/ccisd-teks-educator-mistral7b"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
model = AutoModelForCausalLM.from_pretrained(BASE, device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

messages = [
    {
        "role": "user",
        "content": (
            "Write a lesson objective for a 7th grade science class covering "
            "the flow of energy through an ecosystem."
        ),
    }
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(inputs, max_new_tokens=256)[0]))

Accepting the base model's terms on the Hub is a prerequisite, and the base model needs roughly 15 GB in fp16.

Training data and procedure

Not recorded. No trainer_state.json, training_args.bin, or dataset revision was uploaded with the adapter, so the exact corpus, hyperparameters, epoch count, and hardware cannot be recovered from this repository. Everything in the table above is read off the adapter config; anything beyond it would be a guess and is deliberately omitted.

The TEKS-related datasets published in this namespace are the corpus family this work draws on:

• ccisd-teks-enhanced
• ccisd-teks-alignment
• texas-k12-curriculum-standards-teks
• ccisd-teks-training (deprecated)

Which snapshot of which of these the adapter saw is not documented, so none of them should be cited as the training set.

Limitations

The curriculum data is synthetic. The datasets above are tagged template-generated and synthetic. They were produced programmatically rather than transcribed from official TEKS publications, so standard codes, wording, and grade-band placement in this adapter's output can be plausible and wrong at the same time. That failure mode is the expensive one: fluent, correctly formatted, and unusable without checking.

Verify every standards reference against the official source. The authoritative text is published by the Texas Education Agency. Nothing this adapter produces is a substitute for it.

No evaluation exists. There is no held-out set, no benchmark, and no comparison against the base model. It is unknown whether the adapter improves on Mistral-7B-Instruct-v0.2 for these tasks or degrades it.

Inherited behaviour. All base-model limitations carry over, including fabrication of specifics and English-only competence.

Not for placement or compliance decisions. Do not use it to determine whether instruction meets a standard, to evaluate educators or students, or in any process affecting a student's record.

Affiliation

An independent project. Not affiliated with, endorsed by, or reviewed by the Texas Education Agency or any school district. "CCISD" and "TEKS" appear here as descriptors of the subject matter, not as claims of institutional involvement.

License

Apache 2.0, matching the base model. The associated datasets are CC0-1.0. Use of the adapter also requires accepting the terms of mistralai/Mistral-7B-Instruct-v0.2.