rickyhappy1636/ReasoningwithRicky-7B-v0.1-experimental

ReasoningwithRicky-7B v0.1-experimental

⚠️ This is an early-stage experimental checkpoint, not a production-ready model.

What this is

A LoRA adapter fine-tuned on Qwen/Qwen3-8B, trained entirely on a free-tier Google Colab T4 GPU as a proof-of-concept, across two sessions (resumed from checkpoint after a session interruption).

Honest limitations (please read before using)

• —Training data: only 5,005 examples, 1 epoch. Roughly 1.6% of the originally planned dataset size (~307K examples).
• —Math data is GSM8K only (NuminaMath-CoT excluded from this run due to time constraints).
• —This adapter primarily teaches output format (e.g. <think>...</think> reasoning traces, concise code answers, clean Hindi responses) rather than substantially new capability. Underlying reasoning ability is still mostly that of base Qwen3-8B.
• —Free-tier T4 hardware makes training an 8B-parameter model at meaningful scale impractical in a single day — treat this as a pipeline validation / early checkpoint, not a finished model.
• —No formal benchmark evaluation (GSM8K accuracy, HumanEval, etc.) has been run yet.
• —No safety/toxicity/hallucination testing has been performed.

Training details

Dataset sources

• —Math: openai/gsm8k (MIT)
• —Coding: m-a-p/CodeFeedback-Filtered-Instruction (Apache-2.0)
• —Reasoning: Open-Orca/OpenOrca (MIT) — derived from GPT-4/GPT-3.5 completions, see dataset card for provenance
• —Hindi: soketlabs/bhasha-sft (ayatranslatedflan_cot subset)

Intended use

Research and experimentation only — not recommended for production or high-stakes use.

Roadmap

Future versions aim to scale to the full ~307K example dataset with proper category balancing, multi-epoch training on more capable hardware, and full benchmark evaluation before any stable release.