mdhamidhosen

baamr-math-qwen2.5-1.5b-lora-pilot-577

README

License: apache-2.0

Model Details

Table with columns: Field, Value
Field	Value
Developer / publisher	`mdhamidhosen`
Repository	`mdhamidhosen/baamr-math-qwen2.5-1.5b-lora-pilot-577`
Base model	`Qwen/Qwen2.5-1.5B-Instruct`
Model type	Causal language model with PEFT LoRA adapter
Languages	Bengali (`bn`) and English (`en`)
Domain	Educational mathematics, classes 6–10
Adapter task	`CAUSAL_LM`
License	Apache-2.0

Intended Uses

Suitable pilot uses include:

Bengali or English mathematics question answering;
worked-solution generation for class 6–10 topics;
educational prototyping and qualitative research;
continued fine-tuning and controlled comparison experiments.

Out-of-Scope Uses

Do not use this pilot as an authoritative grading system, an unsupervised source of examination questions, or a substitute for a qualified teacher. It has not been validated for high-stakes assessment, formal proof verification, or domains outside the training scope.

How to Use

Install recent versions of the required packages:

bash
pip install -U "transformers>=4.45,<5" "peft>=0.13" accelerate bitsandbytes torch

Current Transformers/PEFT releases may reject the list-valued extra_special_tokens field in this pilot repository's tokenizer metadata. The robust loading path below uses the official base tokenizer and copies only the two required PEFT adapter files into a clean local directory.

python
import shutil
from pathlib import Path

import torch
from huggingface_hub import hf_hub_download
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

MODEL_ID = "mdhamidhosen/baamr-math-qwen2.5-1.5b-lora-pilot-577"
BASE_MODEL_ID = "Qwen/Qwen2.5-1.5B-Instruct"

compute_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=compute_dtype,
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)

adapter_dir = Path("./baamr_adapter_weights_only")
if adapter_dir.exists():
    shutil.rmtree(adapter_dir)
adapter_dir.mkdir(parents=True)

for filename in ("adapter_config.json", "adapter_model.safetensors"):
    cached_file = hf_hub_download(repo_id=MODEL_ID, filename=filename)
    shutil.copy2(cached_file, adapter_dir / filename)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    quantization_config=quantization_config,
    device_map="auto",
    dtype=compute_dtype,
)
model = PeftModel.from_pretrained(base_model, adapter_dir, is_trainable=False)
model.eval()

messages = [
    {
        "role": "system",
        "content": (
            "You are a bilingual mathematics tutor. Solve the problem clearly, "
            "show the essential steps, and state the final answer."
        ),
    },
    {
        "role": "user",
        "content": "একটি ত্রিভুজের ভূমি ১২ সেমি এবং উচ্চতা ৭ সেমি। ক্ষেত্রফল নির্ণয় কর।",
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(next(model.parameters()).device)

outputs = model.generate(
    **inputs,
    max_new_tokens=384,
    do_sample=False,
)

new_tokens = outputs[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

The same format can be used with an English prompt, for example: A triangle has base 12 cm and height 7 cm. Find its area.

Training Data

The pilot corpus contains 577 paired Bengali–English source records covering 61 chapters across classes 6–10.

Table with columns: Raw split, Records, SFT sequences after BN/EN expansion
Raw split	Records	SFT sequences after BN/EN expansion
Train	457	914
Validation	59	118
Test	61	Held out from training
Total	577	—

Each training record was expanded into separate Bengali and English prompt–completion sequences. Records include questions, worked reasoning, final answers, curriculum metadata, and page-level source references. The corpus passed automated structure, bilingual-parity, duplicate-family, split-leakage, and mathematical-assertion checks. These checks are not equivalent to independent expert review.

The planned 12,000-record corpus is a future target and was not used for this pilot.

Training Procedure

The adapter was trained with supervised fine-tuning (SFT) and 4-bit QLoRA in Google Colab.

Table with columns: Hyperparameter, Value
Hyperparameter	Value
Epochs	3
Maximum sequence length	1,024
Per-device train batch size	2
Gradient accumulation steps	8
Effective batch size (one GPU)	16
Learning rate	1e-4
Scheduler	Cosine
Warmup steps	50
Weight decay	0.01

LoRA Configuration

Table with columns: Parameter, Value
Parameter	Value
Rank (`r`)	16
Alpha	32
Dropout	0.05
Bias	`none`
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, ,

Evaluation

The adapter was evaluated on all 61 held-out raw records in both languages, producing 122 deterministic generations (max_new_tokens=384, do_sample=False, repetition_penalty=1.05). The evaluation dataset SHA-256 is c0f23821e4a8936671993143b3bbe3e61a317139962ada343045c225b770485a.

Table with columns: Held-out metric, Result
Held-out metric	Result
Completion-only test loss	0.501526
Completion-only perplexity	1.6512
Normalized exact final-answer match	33/122 (27.05%)
Bengali exact match	13/61 (21.31%)
English exact match	20/61 (32.79%)
Mean numeric-reference recall	37.83%
Mean final-answer token F1	55.74%
BN/EN generated numeric consistency	8/49 comparable pairs (16.33%)
Automatically flagged for manual review

Exact text matching is conservative because equivalent answers can differ in wording or units. However, manual inspection also found genuine arithmetic errors, bilingual disagreement, and cases where the final answer matched while the written reasoning was defective. Full-response reference-substring matching was 45/122 (36.89%) but is not reported as accuracy, because an incorrect solution can repeat the reference string inside its reasoning.

By question type, exact match was strongest on MCQ (9/12, 75.00%) and weakest on conceptual, multi-step, reasoning, sequence, and short-answer cases (0 exact matches in each category). These automatic metrics are diagnostic and do not establish classroom reliability.

Bias, Risks, and Limitations

The training corpus is small (577 raw records), so topic coverage and linguistic variety remain limited.
The adapter can produce incorrect calculations, invalid reasoning, fabricated facts, or an answer inconsistent with its steps.
Bengali mathematical terminology varies across regions and textbooks; outputs may not match a preferred convention.
Automated dataset checks do not replace review by mathematics teachers and bilingual annotators.
Held-out performance is low and 89/122 generations were automatically flagged for manual review.
Results may be sensitive to prompt wording, decoding settings, library versions, and the base-model revision.
The model should not be trusted for high-stakes educational decisions without independent verification.

Always verify generated solutions, especially geometry arguments, proofs, multi-step algebra, and examination material.

Technical Notes

Adapter format: PEFT LoRA (safetensors)
Adapter file size: 73,911,112 bytes (approximately 70.5 MiB)
Adapter tensors: 392, stored as FP32
adapter_model.safetensors SHA-256: 62f3e9b7ad3413c80505225acd10070ddf003db7faf155883d15d546f23f6c0e
Tokenizer vocabulary size: 151,643
A Qwen-compatible chat template is included in the repository.
The current repository tokenizer metadata has a known compatibility issue with recent Transformers/PEFT releases; use the base-tokenizer loading procedure above until that metadata is replaced.

Environmental Impact

Training was performed in Google Colab. Exact GPU type, runtime duration, energy use, compute region, and carbon emissions were not retained, so an emissions estimate is not reported.

License and Attribution

The base model is published by the Qwen Team under the Apache-2.0 license. Users should review the base-model license and terms as well as the files in this repository before use.

Citation

If you use this adapter, cite the repository:

bibtex
@misc{mdhamidhosen2026baamrpilot577,
  author       = {mdhamidhosen},
  title        = {BaAMR Math Qwen2.5 1.5B LoRA -- Pilot 577},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/mdhamidhosen/baamr-math-qwen2.5-1.5b-lora-pilot-577}}
}

The base model should also be cited according to the instructions in the Qwen2.5-1.5B-Instruct model card.

Acknowledgements

This work uses Qwen2.5, Hugging Face Transformers, PEFT, TRL, Accelerate, and bitsandbytes.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

mdhamidhosen

Model Tree

Base

Qwen/Qwen2.5-1.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model Details

Table with columns: Field, Value
Field	Value
Developer / publisher	`mdhamidhosen`
Repository	`mdhamidhosen/baamr-math-qwen2.5-1.5b-lora-pilot-577`
Base model	`Qwen/Qwen2.5-1.5B-Instruct`
Model type	Causal language model with PEFT LoRA adapter
Languages	Bengali (`bn`) and English (`en`)
Domain	Educational mathematics, classes 6–10
Adapter task	`CAUSAL_LM`
License	Apache-2.0

Intended Uses

Suitable pilot uses include:

Bengali or English mathematics question answering;
worked-solution generation for class 6–10 topics;
educational prototyping and qualitative research;
continued fine-tuning and controlled comparison experiments.

Out-of-Scope Uses

How to Use

Install recent versions of the required packages:

bash
pip install -U "transformers>=4.45,<5" "peft>=0.13" accelerate bitsandbytes torch

python
import shutil
from pathlib import Path

import torch
from huggingface_hub import hf_hub_download
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

MODEL_ID = "mdhamidhosen/baamr-math-qwen2.5-1.5b-lora-pilot-577"
BASE_MODEL_ID = "Qwen/Qwen2.5-1.5B-Instruct"

compute_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=compute_dtype,
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)

adapter_dir = Path("./baamr_adapter_weights_only")
if adapter_dir.exists():
    shutil.rmtree(adapter_dir)
adapter_dir.mkdir(parents=True)

for filename in ("adapter_config.json", "adapter_model.safetensors"):
    cached_file = hf_hub_download(repo_id=MODEL_ID, filename=filename)
    shutil.copy2(cached_file, adapter_dir / filename)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    quantization_config=quantization_config,
    device_map="auto",
    dtype=compute_dtype,
)
model = PeftModel.from_pretrained(base_model, adapter_dir, is_trainable=False)
model.eval()

messages = [
    {
        "role": "system",
        "content": (
            "You are a bilingual mathematics tutor. Solve the problem clearly, "
            "show the essential steps, and state the final answer."
        ),
    },
    {
        "role": "user",
        "content": "একটি ত্রিভুজের ভূমি ১২ সেমি এবং উচ্চতা ৭ সেমি। ক্ষেত্রফল নির্ণয় কর।",
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(next(model.parameters()).device)

outputs = model.generate(
    **inputs,
    max_new_tokens=384,
    do_sample=False,
)

new_tokens = outputs[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

The same format can be used with an English prompt, for example: A triangle has base 12 cm and height 7 cm. Find its area.

Training Data

The pilot corpus contains 577 paired Bengali–English source records covering 61 chapters across classes 6–10.

Table with columns: Raw split, Records, SFT sequences after BN/EN expansion
Raw split	Records	SFT sequences after BN/EN expansion
Train	457	914
Validation	59	118
Test	61	Held out from training
Total	577	—

The planned 12,000-record corpus is a future target and was not used for this pilot.

Training Procedure

The adapter was trained with supervised fine-tuning (SFT) and 4-bit QLoRA in Google Colab.

Table with columns: Hyperparameter, Value
Hyperparameter	Value
Epochs	3
Maximum sequence length	1,024
Per-device train batch size	2
Gradient accumulation steps	8
Effective batch size (one GPU)	16
Learning rate	1e-4
Scheduler	Cosine
Warmup steps	50
Weight decay	0.01

LoRA Configuration

Table with columns: Parameter, Value
Parameter	Value
Rank (`r`)	16
Alpha	32
Dropout	0.05
Bias	`none`
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, ,

Evaluation

Table with columns: Held-out metric, Result
Held-out metric	Result
Completion-only test loss	0.501526
Completion-only perplexity	1.6512
Normalized exact final-answer match	33/122 (27.05%)
Bengali exact match	13/61 (21.31%)
English exact match	20/61 (32.79%)
Mean numeric-reference recall	37.83%
Mean final-answer token F1	55.74%
BN/EN generated numeric consistency	8/49 comparable pairs (16.33%)
Automatically flagged for manual review

Bias, Risks, and Limitations

The training corpus is small (577 raw records), so topic coverage and linguistic variety remain limited.
The adapter can produce incorrect calculations, invalid reasoning, fabricated facts, or an answer inconsistent with its steps.
Bengali mathematical terminology varies across regions and textbooks; outputs may not match a preferred convention.
Automated dataset checks do not replace review by mathematics teachers and bilingual annotators.
Held-out performance is low and 89/122 generations were automatically flagged for manual review.
Results may be sensitive to prompt wording, decoding settings, library versions, and the base-model revision.
The model should not be trusted for high-stakes educational decisions without independent verification.

Always verify generated solutions, especially geometry arguments, proofs, multi-step algebra, and examination material.

Technical Notes

Adapter format: PEFT LoRA (safetensors)
Adapter file size: 73,911,112 bytes (approximately 70.5 MiB)
Adapter tensors: 392, stored as FP32
adapter_model.safetensors SHA-256: 62f3e9b7ad3413c80505225acd10070ddf003db7faf155883d15d546f23f6c0e
Tokenizer vocabulary size: 151,643
A Qwen-compatible chat template is included in the repository.
The current repository tokenizer metadata has a known compatibility issue with recent Transformers/PEFT releases; use the base-tokenizer loading procedure above until that metadata is replaced.

Environmental Impact

Training was performed in Google Colab. Exact GPU type, runtime duration, energy use, compute region, and carbon emissions were not retained, so an emissions estimate is not reported.

License and Attribution

The base model is published by the Qwen Team under the Apache-2.0 license. Users should review the base-model license and terms as well as the files in this repository before use.

Citation

If you use this adapter, cite the repository:

bibtex
@misc{mdhamidhosen2026baamrpilot577,
  author       = {mdhamidhosen},
  title        = {BaAMR Math Qwen2.5 1.5B LoRA -- Pilot 577},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/mdhamidhosen/baamr-math-qwen2.5-1.5b-lora-pilot-577}}
}

The base model should also be cited according to the instructions in the Qwen2.5-1.5B-Instruct model card.

Acknowledgements

This work uses Qwen2.5, Hugging Face Transformers, PEFT, TRL, Accelerate, and bitsandbytes.