After studying this material, you should be able to:

1. Explain what AstaBrief is and why it was built
2. Describe the training pipeline (SFT → DPO) and why each stage matters
3. Understand how data quality filtering improves scientific grounding
4. Evaluate the tradeoffs between speed, cost, and quality in scientific AI systems
5. Recognize the specific challenges of faithful scientific report generation

---

Step-by-Step Teaching

Step 1: The Problem Being Solved

Why existing models weren't enough

Scientists using AI for literature synthesis have unique demands that general chatbots don't handle well:

The specific pain points

Think of it this way: Imagine asking an AI to summarize 50 research papers on a drug treatment. A general model might write something that sounds authoritative but quietly turns a finding from "worked in 40-year-old males" into "works for patients generally." That subtle shift is dangerous in science.

Three core problems existed:

• Speed: Proprietary models took ~178 seconds per report
• Cost: Expensive API calls for every report
• Control: Institutions couldn't run models on their own secure infrastructure

---

Step 2: The Solution Architecture

What AstaBrief actually does

INPUT:  Research question + Retrieved literature excerpts
           ↓
    AstaBrief 8B Model
           ↓
OUTPUT: Cited scientific report (in ONE pass)

Why "one pass" matters

The previous system (Thinking mode) worked like this:

Step 1: Summarize each snippet
Step 2: Cluster summaries
Step 3: Write section 1
Step 4: Write section 2
Step 5: Write section N...

AstaBrief skips all intermediate steps and writes the full report directly. This is why it achieves 3.5× faster generation (51.1 seconds vs 178.5 seconds).

Key insight: More steps ≠ better output. The team found they could skip expensive intermediate stages without sacrificing quality.

---

Step 3: The Training Pipeline

Overview

Base Model (Qwen3-8B)
        ↓
Stage 1: Supervised Fine-Tuning (SFT)
        ↓
Stage 2: Direct Preference Optimization (DPO)
        ↓
AstaBrief 8B

Why start with Qwen3-8B?

• Small enough to be fast and cheap to run
• Large enough to handle complex synthesis
• Open weights = others can reproduce and build on it

---

Step 4: Stage 1 — Supervised Fine-Tuning (SFT)

What is SFT?

SFT teaches the model by showing it examples of correct behavior:

"Here is a question + literature excerpts → Here is what a good report looks like"

The model learns to imitate the target outputs.

Where did the training data come from?

Starting point: Real user queries from Asta (not synthetic benchmarks)

Why real queries matter: Scientists ask differently than regular users. They provide substantial context, multiple constraints, and relationships between concepts — not just keywords.

Filtering pipeline:

Raw user logs (hundreds of thousands)
        ↓
Remove: beta testers, bots, short queries
Remove: non-English, non-scientific, personal info
        ↓
90,000 research-focused queries
        ↓
Generate full reports using Claude 3.5/3.7, o3, o4-mini, GPT-4.1
        ↓
Quality filtering
        ↓
47,000 usable training examples

The critical data quality filters

This is where the team made their most important discovery. They tested four filters:

🔑 Key Lesson #1

Filtering for citation density was the single most powerful improvement.

A report that consistently cites its claims is demonstrating the exact behavior you want: grounding statements in evidence. More complex filter combinations didn't add meaningful gains beyond this simple signal.

Practical implication: You don't need elaborate methods to improve scientific AI. Identifying and filtering for the specific behavior you care about (citation grounding) is more effective than adding complexity.

---

Step 5: Stage 2 — Direct Preference Optimization (DPO)

What is DPO?

Where SFT shows the model "here's a good example," DPO teaches through comparison:

"Given these two reports, Report A is better than Report B — learn why"

How preference pairs were created

Same query → Report A (from Claude pipeline)
           → Report B (from o3, o4-mini, DeepSeek-V3, or DeepSeek-R1)
                ↓
Two judges evaluate: GPT-4.1 AND DeepSeek-R1
                ↓
Only keep pairs where BOTH judges agree
(95% agreement with human preferences)
                ↓
~6,000 clean preference examples

Why require both judges to agree?

Using multiple generators and requiring agreement between two judges:

• Avoids treating any single model as ground truth
• Removes noise from preference data
• Results in cleaner, more reliable training signal

🔑 Key Lesson #2

DPO on top of a strong SFT checkpoint pushed AstaBrief into competitive range with Claude-powered pipelines.

The order matters: get SFT right first, then refine with DPO.

---

Step 6: Why Not Use Reinforcement Learning?

The team considered RL-based training (which has shown strong results in other work) but chose SFT + DPO instead.

The tradeoff: If you use a simpler optimization method, your data quality must compensate. This is why the team invested heavily in data filtering rather than training complexity.

---

Step 7: Evaluating Scientific Faithfulness

The four metrics tracked

1. Rubric Score        → Does the report cover necessary content?
2. Answer Precision    → Is each paragraph relevant to the question?
3. Citation Precision  → Does each citation support its attached claim?
4. Citation Recall     → Are all claims fully supported by citations?

The deeper problem: Scope preservation

Citation support alone isn't enough. A model can cite the right paper and still misrepresent what it says. Common failure modes:

Each step sounds minor. Cumulatively, they can make weak evidence sound like strong consensus.

🔑 Key Lesson #3

A cited sentence can be technically related to its source while still overstating what researchers actually established.

Future evaluations need to test not just whether a claim is cited, but whether the model preserved the evidentiary scope of the original finding.

---

Step 8: Results and Validation

Quantitative results

• AstaBrief was competitive with Claude-powered pipeline on answer and citation quality
• 3.5× faster (51.1s vs 178.5s average)
• Comparable positive feedback rate (84.2% vs 85.2%)

Real-world adoption signals

Among 374 users who tried Fast mode:

29.1%  → Used it on 2+ separate days (returning users)
23%    → Switched to Fast mode and never went back
18%    → Used both modes strategically (~40% Fast mode)
3.67   → Average report threads generated per user

What this tells us: Users aren't just trying it once. They're integrating it into their workflow, which validates the quality-speed tradeoff.

---

Step 9: The Broader Lessons

What generalizes beyond AstaBrief

🔑 Key Lesson #4 (The Meta-Lesson)

Scientific specialization isn't primarily about adding more scientific text to pretraining.

The composition and quality of post-training data — and whether it demonstrates behaviors like grounding and attribution — can materially change how the resulting model performs.

---

Summary: The Complete Picture

PROBLEM
Scientists need fast, grounded, verifiable reports
        ↓
APPROACH
Start with Qwen3-8B (small, open, capable)
        ↓
DATA STRATEGY
Real queries → Generate reports → Filter aggressively
(Citation density filter = most important signal)
        ↓
TRAINING
SFT (learn from good examples)
  + DPO (learn from comparisons)
  = Better than either alone
        ↓
ARCHITECTURE
One-pass generation (skip intermediate steps)
        ↓
RESULT
3.5× faster, competitive quality, open weights
Users can run it locally on sensitive research

---

Self-Check Questions

1. Why did the team choose SFT + DPO over reinforcement learning?
2. What was the single most effective data quality filter, and why does it make intuitive sense?
3. Explain in your own words how a model can cite a paper correctly but still misrepresent the evidence.
4. Why does "one-pass generation" improve speed without necessarily hurting quality?
5. What does the 23% "never switched back" statistic tell us about the quality-speed tradeoff?