Resolve Seq2SeqTrainingArguments Version Compatibility (eval_strategy vs evaluation_strategy)

[purvanshjoshi]

Technical Overview

Recent updates to the Hugging Face transformers library (v4.46+) officially deprecated and subsequently removed the evaluation_strategy parameter in favor of eval_strategy. Conversely, environments pinned to older stable versions of transformers (e.g., v4.22.x) do not recognize eval_strategy and will throw immediate initialization exceptions.

Affected Modules

• File: src/train.py
• Target Class: Seq2SeqTrainingArguments (Line 111-138)

Detailed Traceback / Context

When running on legacy runtimes:

TypeError: __init__() got an unexpected keyword argument 'eval_strategy'

When running on modern runtimes with deprecated keys:

FutureWarning: The 'evaluation_strategy' argument is deprecated and will be removed in a future version, use 'eval_strategy' instead.

Acceptance Criteria

1. Zero-Crash Execution: The training pipeline initializes seamlessly across both legacy (v4.22.0) and modern (v4.48+) environments.
2. Behavioral Parity: Validation evaluation steps occur at the designated save_steps and eval_steps regardless of which underlying key is utilized.
3. No Hardcoding: The script must dynamically inspect the environment rather than using hardcoded try-except blocks that swallow other valid parameters.

Proposed Implementation Approach

Implement a dynamic parameter mapping layer inside the argument dictionary definition prior to class instantiation:

import inspect
from transformers import Seq2SeqTrainingArguments

eval_param = "eval_strategy" if hasattr(Seq2SeqTrainingArguments, "eval_strategy") else "evaluation_strategy"
training_args_dict = {
    # ... other standard arguments
    eval_param: "steps",
}
training_args = Seq2SeqTrainingArguments(**training_args_dict)

Severity & Priority

• Severity: High (Blocks startup on mismatching environments)
• Priority: P1

[Saksham-dev-del]

Hi @purvanshjoshi,

I would like to work on this issue. I will add a dynamic parameter-detection layer in src/train.py (lines 111–138) that inspects the installed transformers version and selects eval_strategy or evaluation_strategy accordingly, rather than hardcoding either key.

I will validate this against both a legacy environment (transformers v4.22.x) and a modern one (v4.48+) to confirm the pipeline initializes cleanly in both cases.

Could you please assign this issue to me? I will submit a PR for review once complete.