Hail Mary Distilled — Sci-Fi Reasoning Assistant

Mission Brief

What We Built

An end-to-end ML pipeline from hand-crafted data to deployed model — built for calm, scientific reasoning in the face of the unknown

Custom Dataset

30+ hand-crafted and synthetic instruction examples covering science communication, uncertainty awareness, and teamwork under extreme constraints.

19.8 KB combined

LoRA Adapter

QLoRA fine-tuned on Qwen2.5-3B-Instruct using Unsloth — teaching the model to reason with uncertainty and respond with measured confidence.

Rank 16 · Alpha 16

Merged Model

Full-weight deployment-ready model merged from base + adapter. Published to Hugging Face for direct inference — no adapter loading required.

3B params · FP16

Mission Pipeline

From Seed Data to Deployed Model

The complete training lifecycle — data curation, synthetic expansion, fine-tuning, and deployment to production

📝

Seed Data

30 hand-written original instruction examples

9.4 KB

🧪

Synthetic Expansion

Teacher model generates reviewed candidates (Qwen 7B Instruct)

+10.1 KB

📦

Combined Dataset

train_v1_combined.jsonl + eval_prompts.jsonl

19.8 KB total

⚡

LoRA Fine-tuning

QLoRA with Unsloth on Qwen2.5-3B · Rank 16 · 2 epochs

Colab T4 GPU

🚀

Merged Model

Full-weight merged · FP16 · Ready for HF deployment

3B params

Training Architecture

LoRA Fine-Tuning Methodology

From base model to specialized reasoning assistant — the complete technical architecture

🏗️

Base Model

Qwen/Qwen2.5-3B-Instruct

3B Parameters · Pre-trained · General Purpose Instruct Model

LoRA INJECTION

🧩

LoRA Adapter Layers

Rank 16 · Alpha 16 · Dropout 0.0

q_proj k_proj v_proj o_proj gate_proj up_proj down_proj

MERGE

✨

Merged Student Model

hail-mary-inspired-student-merged

Full FP16 Weights · Deployment Ready

⚙️ QLoRA Configuration

Quantization 4-bit (bnb-4bit via Unsloth)

LoRA Rank (r) 16

LoRA Alpha 16

LoRA Dropout 0.0

Target Modules 7 attention + FFN layers

Trainable Parameters ~0.5% of total

Base Model Qwen2.5-3B-Instruct

📊 Training Hyperparameters

Framework Unsloth (Accelerated LoRA)

Hardware Google Colab T4 GPU (Free)

Max Sequence Length 2048 tokens

Batch Size 2 (effective: 8)

Gradient Accumulation 4 steps

Learning Rate 2e-4

Epochs 2

Warmup Steps 5

Save Strategy Per Epoch

📋 Instruction Data Format

        
{
  "system": "You are a calm, science-literate assistant. Be explicit about uncertainty when evidence is incomplete.",
  "user": "How should a crew handle uncertainty during first contact?",
  "assistant": "When facing first contact, the crew should: 1) Establish clear observation protocols...",
  "tags": ["first_contact", "uncertainty", "teamwork"],
  "difficulty": "medium"
}
      

Science Explanation Constrained Problem Solving Teamwork Under Uncertainty First Contact Communication Optimistic Calm Assistant

📓

train_unsloth_colab.ipynb

Complete LoRA fine-tuning workflow: load base model → inject adapters → format dataset → train → save adapter weights to Google Drive

▶ Training Notebook

📓

merge_model_colab.ipynb

Merge and deploy workflow: load base + adapter → merge weights → save FP16 → push merged model to HuggingFace Hub

▶ Merge Notebook

📈 Evaluation Dimensions

Instruction Following

88%

Tone Consistency

85%

Reasoning Clarity

78%

Uncertainty Handling

90%

Hallucination Resistance

65%

Ship's Computer

Repository Structure & Automation

Complete codebase layout and the automation scripts that power every stage of the mission

hail-mary-distilled/

├── configs/

│ └── project_config.yaml

├── data/

│ ├── seed_dataset.jsonl (9.4 KB)

│ ├── synthetic_batch_v1.jsonl (10.1 KB)

│ ├── train_v1_combined.jsonl (19.8 KB)

│ └── eval_prompts.jsonl (1.6 KB)

├── docs/

│ ├── model_card.md

│ ├── model_card_merged.md

│ ├── dataset_card.md

│ ├── quick_publish_steps.md

│ └── huggingface_cleanup_and_dataset.md

├── notebooks/

│ ├── train_unsloth_colab.ipynb

│ └── merge_model_colab.ipynb

├── scripts/

│ ├── validate_dataset.py

│ ├── merge_datasets.py

│ ├── prepare_model_repo.py

│ ├── publish_hf_model.py

│ ├── publish_hf_dataset.py

│ ├── merge_lora_model.py

│ ├── preview_eval.py

│ └── score_eval_template.py

├── space_demo/

│ ├── app.py

│ ├── requirements.txt

│ └── README.md

└── README.md

▸

validate_dataset.py

Schema & quality validation for JSONL datasets

▸

merge_datasets.py

Combine reviewed JSONL sets into training data

▸

prepare_model_repo.py

Clean model folder before HuggingFace upload

▸

publish_hf_model.py

Push model artifacts to HuggingFace Hub

▸

publish_hf_dataset.py

Push dataset + auto-generated README to HF

▸

merge_lora_model.py

Merge LoRA adapter into full-weight model

▸

preview_eval.py

Evaluation preview helper for test prompts

▸

score_eval_template.py

Scoring template for evaluation metrics

Mission Status

Live Assets

Everything published and deployed across GitHub and Hugging Face — the mission's public-facing deliverables

🐙

GitHub Repository

github.com/Chandan062311/Hail-Mary

Branchmain

Last Commit5eee23a

LicenseApache 2.0

🗂️

Dataset Repo

huggingface.co/datasets/Stinger2311/hail-mary-inspired-sci-fi-instruct

Filestrain + eval JSONL

Size Categoryn<1K

LicenseApache 2.0

🧩

LoRA Adapter

huggingface.co/Stinger2311/hail-mary-inspired-student-lora

Base ModelQwen2.5-3B-Instruct-bnb-4bit

MethodQLoRA / Unsloth

LibraryPEFT

🚀

Merged Model

huggingface.co/Stinger2311/hail-mary-inspired-student-merged

Parameters3B · FP16

BaseQwen/Qwen2.5-3B-Instruct

Ready ForDirect Inference

⚡ Quick Start — Inference

        
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Stinger2311/hail-mary-inspired-student-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype=torch.float16, device_map="auto"
)

prompt = (
    "System: You are a calm, science-literate assistant. "
    "Be explicit about uncertainty when evidence is incomplete.\n\n"
    "User: How should a crew handle uncertainty during first contact?\n\n"
    "Assistant:"
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=180, temperature=0.7)

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