Samalas/govcon-qwen-7b-lifecycle

GovCon Qwen-7B Lifecycle Model

A fine-tuned Qwen2.5-7B model specialized for Government Contracting (GovCon) lifecycle management.

Overview

This model is optimized for handling all 5 phases of the government contracting lifecycle:

1. Pre-Solicitation - RFI analysis, competitive intelligence, capability gap assessment
2. RFP Compliance - FAR Part 15 parsing, compliance matrix generation, ambiguity detection
3. Debrief & Protest - Weakness analysis, FAR 15.506 protest argument generation
4. Post-Award Execution - CDRL tracking, CPARS monitoring, invoice reconciliation
5. Recompete Radar - Win probability modeling, CPARS trend analysis, competitive positioning

Model Details

• Base Model: mlx-community/Qwen2.5-7B-Instruct-4bit
• Adapter Type: LoRA (Low-Rank Adaptation)
  • Rank: 64
  • Alpha: 32
  • Dropout: 0.05
• Training Data: 69 ChatML-formatted instruction-response pairs
  • Training set: 55 samples
  • Validation set: 14 samples
  • Coverage: 5 lifecycle phases (13-16 samples per phase)
• Training Config:
  • Batch size: 1 (M2 Pro optimized)
  • Iterations: 600 (~11x over full dataset)
  • Learning rate: 1e-5
  • Seed: 42
• Hardware: Apple M2 Pro (16GB unified memory, MLX Metal)
• Framework: MLX (machine learning framework optimized for Apple Silicon)

Intended Use

This model is designed for organizations and consultants involved in government contracting to:

Pre-Solicitation Phase

• Automatically extract requirements from RFI/Sources Sought notices
• Generate competitive intelligence summaries
• Identify incumbent weaknesses and entry strategies
• Assess capability maturity levels

RFP Compliance Phase

• Parse complex FAR Part 15 solicitation documents (Sections A-M)
• Map RFP requirements to proposal sections
• Flag ambiguous language requiring Government clarification
• Generate compliance checklists and attestation requirements

Debrief & Protest Phase

• Analyze debrief feedback for logical inconsistencies
• Draft FAR 15.506 protest arguments
• Score evaluation methodology for fairness
• Recommend remedies and escalation strategies

Post-Award Execution Phase

• Automatically generate CDRL (Contract Data Requirements List) tracking matrices
• Monitor CPARS (Contractor Performance Assessment Rating System) trends
• Reconcile invoices against CLINs (Contract Line Items)
• Flag cost/schedule variance requiring corrective action

Recompete Radar Phase

• Model win probabilities using Bayesian scoring
• Forecast CPARS improvement opportunities
• Evaluate teaming and partnership strategies
• Develop competitive differentiation narratives

Prompt Format

Use ChatML format (compatible with Qwen2.5-Instruct):

{
  "messages": [
    {"role": "user", "content": "Your GovCon instruction/question"},
    {"role": "assistant", "content": "Model generates structured response (often JSON)"}
  ]
}

Example Prompts

Pre-Solicitation: Extract RFI Requirements

User: "Extract key capability requirements from this RFI:\n\n'Government seeks vendors capable of: (1) Real-time data processing at 10K events/sec, (2) NIST 800-53 security compliance, (3) Support for 500+ concurrent users, (4) Integration with legacy COBOL systems.'"

Expected Output: JSON with phase, extracted_requirements (array with req_id, type, text, priority), gaps_to_address, recommended_response_strategy

RFP Compliance: Parse Section L Quality Requirement

User: "Parse this RFP Section L quality requirement and identify compliance gaps:\n\n'System must achieve 99.95% availability (measured monthly). Automated testing must cover 90%+ of codebase. Defect SLA: P1 (4 hours), P2 (2 days), P3 (5 days). Monthly performance reports required by 5th business day.'"

Expected Output: JSON with phase, quality_metrics, reporting_requirements, penalties, incident_response_sla, compliance_strategy

Debrief & Protest: Analyze Losing Bid

User: "Analyze this debrief feedback:\n\n'Your technical score: 82/100 (winner: 75). Your cost: $5.2M (winner: $4.1M). Past performance: You 70/100, Winner 80/100. You lost despite technical advantage. Is this protest-worthy?'"

Expected Output: JSON with phase, your_competitive_position, evaluation_narrative, protest_argument (claim, premise, question), recompete_strategy

Post-Award: Create CDRL Tracking Matrix

User: "Convert this schedule into a CDRL milestone tracking sheet:\n\n'M1: Requirements finalization (30 days). M2: Architecture design (45 days after M1). M3: Core module development (60 days after M2). M4: Integration & testing (45 days after M3). M5: UAT support (30 days after M4). M6: Go-live & knowledge transfer (14 days after M5).'"

Expected Output: JSON with phase, milestone_schedule (array of milestones), total_project_duration, payment_schedule

Recompete: Win Probability Analysis

User: "Analyze recompete strategy:\n\n'Current contract: $10M/5yr, expires 2027-12-31. Incumbent CPARS: 3.2/4.0. Your firm: 3.4/4.0 past performance. Budget RFP June 2027, proposal due Sept 2027. What is your win probability?'"

Expected Output: JSON with phase, contract_details, recompete_timeline, competitive_analysis, critical_path_tasks, win_probability, recommended_differentiators

Performance Notes

• Output Format: Model primarily generates structured JSON responses for programmatic consumption
• Structured Data: Responses include phase context, analysis details, and actionable recommendations
• Accuracy: Trained on 69 representative GovCon scenarios covering ~11x data augmentation
• Hallucination Risk: Model may invent details in underrepresented scenarios; validate against actual solicitation documents
• Best For: GovCon proposal teams, capture managers, bid/no-bid analysts, post-award program managers

Limitations

• Training data is synthetic/representative; may not cover all unique Government processes
• Model is 7B parameters; may struggle with very long (>2000 token) documents
• Specialized vocabulary; general-purpose LLM capabilities remain intact but not optimized
• Not intended as substitute for Government compliance or legal review
• Trained only on English language contracts

How to Use

Install MLX and Load the Model

pip install mlx mlx-lm

# Python inference
from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/Qwen2.5-7B-Instruct-4bit",
    adapter_path="govcon-qwen-7b-lifecycle"  # path to this model's adapter
)

# Generate response
response = generate(
    model,
    tokenizer,
    prompt="Your GovCon question here",
    max_tokens=1000,
    temperature=0.7
)
print(response)

Command-Line Usage (if fused model)

mlx_lm.generate --model govcon-qwen-7b-lifecycle --prompt "Your GovCon instruction"

Citation

If you use this model, please cite:

@misc{govcon-qwen-7b-lifecycle,
  title={GovCon Qwen-7B Lifecycle Model},
  author={Your Name},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/YOUR_USERNAME/govcon-qwen-7b-lifecycle}}
}

License

This model inherits the license of the base Qwen2.5-7B model. See Qwen/Qwen2.5-7B for details.

Disclaimer

This model is provided for educational and research purposes. Users are responsible for:

• Validating model outputs against official Government solicitation documents
• Ensuring compliance with applicable FAR regulations
• Obtaining appropriate legal and compliance review
• Using outputs only in authorized Government contracting contexts

The model does not provide legal advice and should not be relied upon as a substitute for professional Government contracts counsel.

---

Built with: MLX Framework | Qwen2.5-7B-Instruct | LoRA Fine-tuning Optimized for: Apple Silicon (M-series) | macOS