Revise model card: correct numbers and license, remove unmeasured claims, add limitations

README.md

@@ -1,669 +1,174 @@
 ---
-license: mit
+license: llama3.1
 language:
 - en
 tags:
 - agriculture
 - business-intelligence
-- b2b
-- fine-tuned
 - llama
 - llama-3.1
 - lora
 - peft
 - knowledge-distillation
-- multi-teacher
-- claude
-- gpt-4
+- sequence-level-distillation
+- synthetic-data
+- experiment
 library_name: peft
 base_model: meta-llama/Llama-3.1-8B-Instruct
-datasets:
-- custom
 pipeline_tag: text-generation
-model-index:
-- name: llama-agri-b2b-intelligence
-  results:
-  - task:
-      type: text-generation
-      name: Agricultural B2B Analysis
-    dataset:
-      type: custom
-      name: Agricultural B2B Intelligence Dataset
-    metrics:
-    - type: train_loss
-      value: 0.70
-      name: Final Training Loss
-    - type: eval_loss
-      value: 0.78
-      name: Final Evaluation Loss
-    - type: training_time
-      value: 55
-      name: Training Hours
 ---
 
-# Agricultural B2B Intelligence Model
+# Llama 3.1 8B — Agricultural B2B Intelligence (LoRA adapter, experiment)
 
-<div align="center">
+**Built with Llama.**
 
-![Version](https://img.shields.io/badge/version-1.0.0-blue)
-![License](https://img.shields.io/badge/license-MIT-green)
-![Model Size](https://img.shields.io/badge/parameters-8B-orange)
-![Training Hours](https://img.shields.io/badge/training-55%20hours-purple)
+> **Personal experiment** (December 2025). This LoRA adapter for Llama-3.1-8B-Instruct was trained
+> on 10,000 agricultural business-analysis answers produced by two commercial LLMs. It is a
+> study of **sequence-level knowledge distillation**: supervised fine-tuning on teacher outputs.
+> So far it has been evaluated only by training and validation loss. **No task benchmark has been run yet.**
 
-**A production-grade language model fine-tuned for agricultural business intelligence, built using multi-teacher knowledge distillation from Claude Sonnet 4 and GPT-4.1**
+Merged-weights version: [llama-agri-b2b-intelligence-merged](https://huggingface.co/sarathi-balakrishnan/llama-agri-b2b-intelligence-merged)
 
-[Model Card](#model-details) | [Usage](#usage) | [Training Journey](#the-journey) | [Metrics](#training-metrics) | [API](#api-usage)
+## Summary
 
-</div>
+| | |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B-Instruct` (8.03B) |
+| Method | LoRA SFT on teacher-generated responses (sequence-level distillation) |
+| Teacher outputs | 5,007 from Claude Sonnet 4, 4,993 from GPT-4.1 (10,000 total, ~1,150 tokens avg) |
+| Split | 8,500 train / 1,000 validation / 500 test (test set not yet used) |
+| LoRA | r = 128, α = 256, dropout 0.05, all 7 projections in every layer |
+| Trainable parameters | **335,544,320** (≈4.18% of 8.03B) |
+| Adapter file | 1.34 GB (`adapter_model.safetensors`, FP32) |
+| Hardware / time | 1× NVIDIA DGX Spark (GB10), ~55 h for 3 epochs (798 steps, ~4 min/step, ~45 GB memory) |
+| Evaluation | Loss only. Benchmark **pending** |
 
----
-
-## Executive Summary
-
-This model represents a breakthrough in agricultural AI - distilling the combined knowledge of two frontier AI models (Claude Sonnet 4 and GPT-4.1) into a compact, deployable 8B parameter model. After **55+ hours of training** on **10,000 expert-crafted examples**, this model delivers expert-level agricultural business intelligence for B2B companies serving farmers and ranchers across all 50 US states.
-
-```
-┌─────────────────────────────────────────────────────────────────────────────┐
-│                    MULTI-TEACHER KNOWLEDGE DISTILLATION                      │
-├─────────────────────────────────────────────────────────────────────────────┤
-│                                                                              │
-│    ┌──────────────────────┐            ┌──────────────────────┐             │
-│    │    TEACHER 1         │            │    TEACHER 2         │             │
-│    │    Claude Sonnet 4   │            │    GPT-4.1           │             │
-│    │    5,007 responses   │            │    4,993 responses   │             │
-│    │    ~$100 API cost    │            │    ~$80 API cost     │             │
-│    └──────────┬───────────┘            └──────────┬───────────┘             │
-│               │                                   │                          │
-│               └─────────────┬─────────────────────┘                          │
-│                             │                                                │
-│                             ▼                                                │
-│               ┌─────────────────────────┐                                   │
-│               │     STUDENT MODEL       │                                   │
-│               │  Llama 3.1 8B Instruct  │                                   │
-│               │     LoRA Fine-tuned     │                                   │
-│               │    55 hours training    │                                   │
-│               └─────────────────────────┘                                   │
-│                                                                              │
-└─────────────────────────────────────────────────────────────────────────────┘
-```
-
----
-
-## The Journey: From Phi-3 to Llama 3.1
-
-### The Evolution of Our Student Model Selection
-
-This project went through several iterations before arriving at the optimal architecture. Here's the complete story:
-
-#### Phase 1: Phi-3 Mini (Abandoned)
-**Initial Consideration:** Microsoft's Phi-3 Mini (3.8B parameters)
-- **Pros:** Small, fast, efficient
-- **Cons:** Limited context window, struggled with complex agricultural terminology
-- **Decision:** Too small for the domain complexity required
-
-#### Phase 2: Qwen 2.5 7B (Temporary)
-**Second Attempt:** Alibaba's Qwen 2.5-7B-Instruct
-- **Pros:** Open weights, good multilingual support, strong reasoning
-- **Cons:** Less optimized for English-only use case, licensing considerations for commercial use
-- **Status:** Used temporarily while awaiting Llama access approval
-- **Why we moved on:** Meta's Llama offered better ecosystem support and commercial licensing
-
-#### Phase 3: Llama 3.1 8B Instruct (Final Selection)
-**Final Choice:** Meta's Llama 3.1-8B-Instruct
-- **Pros:**
-  - Excellent instruction following
-  - Strong reasoning capabilities
-  - Permissive license for commercial use
-  - Large community and ecosystem
-  - Optimized for chat/instruct use cases
-  - Native support for long contexts (128K)
-- **Cons:** Gated model requiring license approval
-- **Decision:** Best balance of capability, licensing, and community support
-
-### Why Multi-Teacher Distillation?
-
-Traditional fine-tuning uses a single data source. We pioneered a **multi-teacher approach**:
-
-| Aspect | Single Teacher | Multi-Teacher (Our Approach) |
-|--------|----------------|------------------------------|
-| Diversity | Limited perspective | Complementary viewpoints |
-| Robustness | May inherit biases | Cross-validated knowledge |
-| Coverage | Gaps in knowledge | Comprehensive coverage |
-| Quality | Single style | Best of both worlds |
-
-**Claude Sonnet 4** excels at:
-- Structured, methodical analysis
-- Nuanced risk assessment
-- Detailed data interpretation
+## How the data was made
 
-**GPT-4.1** excels at:
-- Creative market insights
-- Trend identification
-- Actionable recommendations
+1. I built a knowledge base and a query generator for US agricultural regions and B2B sectors
+   (crop insurance, equipment, seed, fertilizer, crop protection, irrigation, lending, land brokerage).
+2. I generated 10,000 queries. The mix: county 30%, zip code 15%, market 10%, state 10%, risk 8%,
+   B2B marketing 8%, predictions 6%, historical trends 6%, commodity 4%, comparative 3%.
+3. Each query was sent once to **one** teacher through the providers' APIs. That gave 5,007 Claude Sonnet 4
+   responses (~$100 of API usage) and 4,993 GPT-4.1 responses (~$80).
+4. The student was fine-tuned with standard next-token cross-entropy on those responses. No teacher logits
+   were used (they aren't available from these APIs), so this is sequence-level KD, not logit/KL distillation.
 
-By combining both, our student model inherits the **strengths of each teacher** while mitigating individual weaknesses.
+**About the teacher data.** The training targets are outputs from third-party models. Anyone using this
+adapter or its data is also subject to the terms of the providers whose models generated those
+outputs (Anthropic and OpenAI), in addition to the Llama 3.1 Community License. Please review those terms
+for your use case. The responses weren't fact-checked, so any statistics they contain may be wrong.
 
----
-
-## Training Infrastructure
-
-### Hardware Configuration
-
-```
-┌─────────────────────────────────────────────────────────┐
-│              NVIDIA DGX SPARK WORKSTATION               │
-├─────────────────────────────────────────────────────────┤
-│  GPU:     NVIDIA GB10 (Blackwell Architecture)         │
-│  VRAM:    128 GB Unified Memory                         │
-│  RAM:     128 GB System Memory                          │
-│  Storage: 4 TB NVMe SSD                                 │
-│  OS:      Ubuntu Linux                                  │
-└─────────────────────────────────────────────────────────┘
-```
-
-### Why Blackwell GPU?
-
-The NVIDIA Blackwell GB10 represents the cutting edge of AI training hardware:
-- **BF16 Native Support:** Optimal precision for LLM training
-- **Unified Memory:** 128GB allows full model + gradients in memory
-- **Tensor Cores:** 5th generation for maximum throughput
-- **Energy Efficiency:** Lower power consumption than previous generations
-
----
+## Training metrics
 
-## Complete Pipeline Timeline
+| Metric | Epoch 1 | Epoch 2 | Epoch 3 |
+|---|---|---|---|
+| Training loss | 0.89 | 0.77 | ~0.70 |
+| Validation loss | 0.87 | 0.82 | ~0.78 |
+| Cumulative steps | 266 | 532 | 798 |
+| Wall-clock | 18.5 h | 18.5 h | 18 h |
+| Learning rate (cosine) | 5e-5 → 4.1e-5 | 4.1e-5 → 1.6e-5 | 1.6e-5 → 0 |
 
-### Total Project Duration: ~70 Hours
-
-```
-Phase                          Duration    Status
-─────────────────────────────────────────────────
-1. Environment Setup           30 min      ✅ Complete
-2. Knowledge Base Creation     45 min      ✅ Complete
-3. Query Generation            20 min      ✅ Complete
-4. Teacher Response Gen        8 hours     ✅ Complete
-   ├─ Claude Sonnet 4          4.5 hrs     (5,007 responses)
-   └─ GPT-4.1                  3.5 hrs     (4,993 responses)
-5. Dataset Preparation         15 min      ✅ Complete
-6. Model Fine-tuning           55 hours    ✅ Complete
-   ├─ Epoch 1                  18.5 hrs
-   ├─ Epoch 2                  18.5 hrs
-   └─ Epoch 3                  18.0 hrs
-7. Benchmarking                30 min      ⏳ Pending
-8. HuggingFace Upload          10 min      ⏳ Pending
-─────────────────────────────────────────────────
-TOTAL                          ~70 hours
-```
-
----
-
-## Training Metrics
-
-### Loss Progression Across Epochs
-
-```
-Loss
-│
-1.8 ┤ ●
-    │  ╲
-1.6 ┤   ╲
-    │    ╲
-1.4 ┤     ╲
-    │      ╲
-1.2 ┤       ●
-    │        ╲
-1.0 ┤         ●──●
-    │             ╲
-0.8 ┤              ●──●──●──●──●──●──●──●
-    │                    EPOCH 1 │ EPOCH 2 │ EPOCH 3
-0.6 ┤
-    │
-0.4 ┼────┬────┬────┬────┬────┬────┬────┬────┬────
-    0   100  200  300  400  500  600  700  800
-                        Steps
-```
-
-### Detailed Epoch Metrics
-
-| Metric | Epoch 1 | Epoch 2 | Epoch 3 | Improvement |
-|--------|---------|---------|---------|-------------|
-| **Training Loss** | 0.89 | 0.77 | ~0.70 | **-21%** |
-| **Eval Loss** | 0.87 | 0.82 | ~0.78 | **-10%** |
-| **Steps** | 266 | 532 | 798 | - |
-| **Duration** | 18.5 hrs | 18.5 hrs | 18 hrs | - |
-| **Learning Rate** | 5e-5 → 4.1e-5 | 4.1e-5 → 1.6e-5 | 1.6e-5 → 0 | Cosine decay |
-
-### Training Configuration
+By epoch 3, training loss (~0.70) and validation loss (~0.78) are starting to separate. A 4th epoch would
+probably overfit.
 
 ```python
-# LoRA Configuration
-LORA_CONFIG = {
-    "r": 128,                    # High rank for complex domain
-    "lora_alpha": 256,           # 2x rank for stable training
-    "target_modules": [
-        "q_proj", "k_proj", "v_proj", "o_proj",
-        "gate_proj", "up_proj", "down_proj"
-    ],
-    "lora_dropout": 0.05,
-    "bias": "none",
-    "task_type": "CAUSAL_LM"
-}
-
-# Training Configuration
-TRAINING_CONFIG = {
-    "num_train_epochs": 3,
-    "per_device_train_batch_size": 2,
-    "gradient_accumulation_steps": 16,  # Effective batch = 32
-    "learning_rate": 5e-5,
-    "warmup_ratio": 0.1,
-    "lr_scheduler_type": "cosine",
-    "bf16": True,
-    "gradient_checkpointing": True,
-    "max_seq_length": 4096
-}
-```
-
-### Parameter Efficiency
-
-| Metric | Value |
-|--------|-------|
-| Base Model Parameters | 8,030,261,248 (8.03B) |
-| LoRA Trainable Parameters | 167,772,160 (168M) |
-| **Trainable Ratio** | **2.09%** |
-| Memory Usage | ~45 GB VRAM |
-| Training Throughput | ~4 min/step |
-
----
-
-## Dataset Composition
-
-### Knowledge Base Coverage
-
-```
-┌─────────────────────────────────────────────────────────┐
-│              COMPREHENSIVE US AGRICULTURAL DATA          │
-├─────────────────────────────────────────────────────────┤
-│                                                          │
-│  🗺️  Geographic Coverage                                │
-│  ├── States:        50 (All US states)                  │
-│  ├── Counties:      3,142 (99.9% coverage)              │
-│  ├── Zipcodes:      2,000 (Key agricultural areas)      │
-│  └── Regions:       All USDA Farm Resource Regions      │
-│                                                          │
-│  🌾  Agricultural Data                                   │
-│  ├── Crops:         12 major commodities                │
-│  ├── Livestock:     8 categories                        │
-│  └── Markets:       15 commodity markets                │
-│                                                          │
-│  🏢  Industry Coverage                                   │
-│  └── B2B Sectors:   8 specialized industries            │
-│                                                          │
-└─────────────────────────────────────────────────────────┘
-```
-
-### Query Distribution (10,000 Total)
-
-| Category | Count | Percentage | Description |
-|----------|-------|------------|-------------|
-| County Intelligence | 3,000 | 30% | Deep county-level analysis |
-| Zipcode Analysis | 1,500 | 15% | Granular local insights |
-| Market Intelligence | 1,000 | 10% | Market trends and opportunities |
-| State Analysis | 1,000 | 10% | State-wide agricultural overview |
-| Risk Assessment | 800 | 8% | Risk evaluation and mitigation |
-| B2B Marketing | 800 | 8% | Go-to-market strategies |
-| Predictions | 600 | 6% | Future trend forecasting |
-| Historical Trends | 600 | 6% | Historical data analysis |
-| Commodity Analysis | 400 | 4% | Commodity-specific insights |
-| Comparative Analysis | 300 | 3% | Cross-region comparisons |
-
-### Teacher Response Statistics
-
-| Teacher | Responses | Avg Length | Avg Time | Total Cost |
-|---------|-----------|------------|----------|------------|
-| Claude Sonnet 4 | 5,007 | ~1,200 tokens | 4.2s | ~$100 |
-| GPT-4.1 | 4,993 | ~1,100 tokens | 2.8s | ~$80 |
-| **Combined** | **10,000** | ~1,150 tokens | 3.5s | **~$180** |
-
-### Data Split
-
-```
-Training:    8,500 examples (85%)  ████████████████░░
-Validation:  1,000 examples (10%)  ██░░░░░░░░░░░░░░░░
-Test:          500 examples (5%)   █░░░░░░░░░░░░░░░░░
-─────────────────────────────────────────────────────
-Total:      10,000 examples
-```
-
----
-
-## Target Industries & Use Cases
-
-### Supported B2B Sectors
-
-| Industry | Primary Use Cases | Key Metrics Provided |
-|----------|-------------------|----------------------|
-| **Crop Insurance** | Risk assessment, premium pricing, loss prediction | Historical loss ratios, weather risk scores, yield variability |
-| **Farm Equipment** | Market sizing, dealer network, territory planning | Equipment penetration, farm size distribution, mechanization rates |
-| **Seed & Genetics** | Variety placement, market penetration, climate zones | Seed market share, variety performance, adoption curves |
-| **Fertilizer & Soil** | Demand forecasting, logistics, pricing | Soil types, nutrient needs, application rates |
-| **Pesticides** | Application timing, resistance patterns, compliance | Pest pressure maps, resistance tracking, regulatory status |
-| **Irrigation** | Water management, system sizing, ROI analysis | Water availability, irrigation penetration, efficiency metrics |
-| **Agricultural Lending** | Farm credit risk, land valuation, cash flow | Debt ratios, land values, income stability |
-| **Land Brokerage** | Parcel analysis, comparable sales, investment returns | Price per acre trends, transaction volumes, cap rates |
-
----
-
-## Model Capabilities
-
-### What This Model Can Do
-
-✅ **Geographic Analysis**
-- State-level agricultural overviews
-- County-level deep dives
-- Zipcode-level granular insights
-- Cross-region comparisons
-
-✅ **Market Intelligence**
-- Market entry analysis
-- Competitive landscape mapping
-- Opportunity identification
-- Risk factor assessment
-
-✅ **Business Strategy**
-- Go-to-market recommendations
-- Territory planning
-- Customer segmentation
-- Pricing strategy insights
-
-✅ **Data Synthesis**
-- USDA data interpretation
-- Trend analysis
-- Predictive insights
-- Historical pattern recognition
-
-### Output Format
-
-The model generates structured analysis following this format:
-
-```markdown
-## Executive Summary
-[2-3 sentence high-level overview]
-
-## Geographic/Market Profile
-[Key statistics and characteristics]
-
-## Analysis
-[Detailed data-driven insights with specific metrics]
-
-## Opportunities
-[Specific actionable opportunities ranked by potential]
-
-## Risk Factors
-[Challenges, constraints, and mitigation strategies]
-
-## Recommendations
-[Prioritized action items with implementation guidance]
-
-## Confidence Level
-[High/Medium/Low with reasoning and data quality notes]
-
-## Data Sources
-[Referenced USDA and industry sources]
-```
-
----
+LORA_CONFIG = dict(r=128, lora_alpha=256, lora_dropout=0.05, bias="none", task_type="CAUSAL_LM",
+                   target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
+                                   "gate_proj", "up_proj", "down_proj"])
+TRAINING_CONFIG = dict(num_train_epochs=3, per_device_train_batch_size=2,
+                       gradient_accumulation_steps=16,   # effective batch 32
+                       learning_rate=5e-5, warmup_ratio=0.1, lr_scheduler_type="cosine",
+                       bf16=True, gradient_checkpointing=True, max_seq_length=4096)
+```
+
+**Parameter count check.** Each layer adds r·(d_in + d_out) parameters per projection. For Llama-3.1-8B
+(hidden 4096, KV dim 1024, MLP 14336) at r = 128 that gives 10,485,760 per layer. Across 32 layers that is
+**335,544,320**. An earlier version of this card said 167,772,160 (2.09%), which is the r = 64 figure.
+The FP32 adapter file size (1,342,238,560 bytes ≈ 335.5M × 4 bytes) confirms the corrected number.
+
+## What's missing / lessons learned
+
+- **Loss isn't quality.** Lower validation loss shows the student imitates the teachers' style. It
+  doesn't show the answers are correct or useful. The planned evaluation is to run the 500-example test
+  split through base Llama-3.1-8B-Instruct, this adapter, and both teachers, then score with an LLM judge
+  using a fixed rubric (factuality, structure, actionability) plus a human spot check.
+- **Rank 128 was probably more than needed.** Lower ranks (16–64) are common for style and
+  format adaptation. A rank sweep would show whether the 1.34 GB adapter could be much smaller.
+- **I didn't track which teacher the student follows.** A per-teacher validation loss would show
+  whether mixing two teachers helps or just averages their styles.
+- **Throughput.** About 4 minutes per optimizer step (32 sequences × up to 4,096 tokens) on a single GB10.
+  Sequence packing and a smaller rank would cut the 55 h substantially.
 
 ## Usage
 
-### Quick Start
-
 ```python
 from transformers import AutoTokenizer, AutoModelForCausalLM
 from peft import PeftModel
 import torch
 
-# Load model
 base_model_name = "meta-llama/Llama-3.1-8B-Instruct"
 adapter_name = "sarathi-balakrishnan/llama-agri-b2b-intelligence"
 
 tokenizer = AutoTokenizer.from_pretrained(base_model_name)
 base_model = AutoModelForCausalLM.from_pretrained(
-    base_model_name,
-    torch_dtype=torch.bfloat16,
-    device_map="auto"
+    base_model_name, torch_dtype=torch.bfloat16, device_map="auto"
 )
 model = PeftModel.from_pretrained(base_model, adapter_name)
 
-# Generate analysis
 messages = [
-    {
-        "role": "system",
-        "content": "You are an expert agricultural business intelligence analyst."
-    },
-    {
-        "role": "user",
-        "content": "Analyze Iowa's corn market for a seed company evaluating expansion."
-    }
+    {"role": "system", "content": "You are an agricultural business intelligence analyst."},
+    {"role": "user", "content": "Analyze Iowa's corn market for a seed company evaluating expansion."},
 ]
-
 text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
 inputs = tokenizer(text, return_tensors="pt").to(model.device)
 
-outputs = model.generate(
-    inputs["input_ids"],
-    max_new_tokens=1500,
-    temperature=0.7,
-    do_sample=True
-)
-
-response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
-print(response)
+outputs = model.generate(**inputs, max_new_tokens=1500, temperature=0.7, top_p=0.9,
+                         repetition_penalty=1.1, do_sample=True)
+print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
 ```
 
-### Merging LoRA Weights (Optional)
-
-For faster inference, merge the LoRA adapters:
+Optional merge:
 
 ```python
-# Merge and save
-merged_model = model.merge_and_unload()
-merged_model.save_pretrained("./llama-agri-merged")
+merged = model.merge_and_unload()
+merged.save_pretrained("./llama-agri-merged")
 tokenizer.save_pretrained("./llama-agri-merged")
 ```
 
-### Recommended Generation Parameters
-
-```python
-generation_config = {
-    "max_new_tokens": 1500,      # Sufficient for detailed analysis
-    "temperature": 0.7,          # Balanced creativity/consistency
-    "top_p": 0.9,                # Nucleus sampling
-    "repetition_penalty": 1.1,   # Reduce repetition
-    "do_sample": True
-}
-```
-
----
+### Output format the model was trained to follow
 
-## Example Outputs
-
-### Example 1: County-Level Analysis
-
-**Query:** "Analyze agricultural opportunities in Fresno County, California for a precision agriculture company."
-
-**Response:**
-```markdown
-## Executive Summary
-Fresno County represents one of the nation's most valuable agricultural markets,
-with over $7.5 billion in annual agricultural output. The combination of diverse
-crop production, large farm operations, and water management challenges creates
-significant opportunities for precision agriculture solutions.
-
-## Geographic Profile
-- **Location:** Central San Joaquin Valley, California
-- **Agricultural Acres:** 1.9 million acres
-- **Top Crops:** Almonds, grapes, tomatoes, cotton, citrus
-- **Farm Count:** ~6,000 operations
-- **Average Farm Size:** 315 acres
-
-## Analysis
-Fresno County leads California in agricultural production value. Key factors:
-
-1. **Crop Diversity:** 350+ commodities grown commercially
-2. **Irrigation Dependency:** 95% of farmland irrigated
-3. **Water Stress:** Ongoing groundwater sustainability concerns
-4. **Technology Adoption:** 35% precision ag penetration (below potential)
-
-## Opportunities
-1. **Water Management Solutions** (High Priority)
-   - Soil moisture monitoring: $15M addressable market
-   - Variable rate irrigation: Growing 25% annually
-
-2. **Specialty Crop Analytics**
-   - Tree nut optimization: 500,000+ almond acres
-   - Vineyard management: 180,000 grape acres
-
-## Risk Factors
-- Water availability regulations (SGMA compliance)
-- Labor cost pressures driving automation
-- Climate variability affecting crop selection
-
-## Recommendations
-1. Partner with irrigation districts for pilot programs
-2. Focus initial sales on almond and grape operations (highest ROI)
-3. Develop Spanish-language support (40% Hispanic farm operators)
-
-## Confidence Level
-**High** - Based on USDA Census of Agriculture, California Department of
-Food and Agriculture data, and established market trends.
-```
-
----
+Executive Summary → Geographic/Market Profile → Analysis → Opportunities → Risk Factors →
+Recommendations → Confidence Level → Data Sources.
 
 ## Limitations
 
-### Known Limitations
-
-1. **Geographic Scope:** Focused exclusively on US agricultural markets
-2. **Data Currency:** Training data reflects knowledge up to early 2025
-3. **Synthetic Training:** Responses generated by AI teachers, not human experts
-4. **Specificity:** May lack hyperlocal details for small rural communities
-5. **Numerical Precision:** Statistics should be verified against official sources
+- **Not benchmarked.** See above.
+- **Synthetic, unverified targets.** The responses were written by LLMs and not checked by domain experts.
+  Figures such as acreage, market sizes, and percentages can be invented. Verify them against USDA NASS/ERS or other primary sources.
+- US agriculture only. English only. Trained on 4,096-token sequences.
+- Not for regulatory, lending, insurance, investment, or trading decisions.
 
-### When NOT to Use This Model
+## Files
 
-- For official regulatory compliance decisions
-- As sole source for financial investment decisions
-- For real-time commodity trading signals
-- To replace professional agricultural consultants for high-stakes decisions
-
-### Recommended Usage
-
-This model is best used as:
-- A starting point for market research
-- A tool for generating initial analysis frameworks
-- A complement to (not replacement for) professional expertise
-- A rapid prototyping tool for agricultural B2B applications
-
----
+`adapter_config.json`, `adapter_model.safetensors` (1.34 GB), tokenizer files.
 
-## Technical Specifications
-
-### Model Architecture
-
-| Component | Specification |
-|-----------|---------------|
-| Base Model | Llama 3.1 8B Instruct |
-| Architecture | Transformer (decoder-only) |
-| Parameters | 8.03B (base) + 168M (LoRA) |
-| Context Length | 4,096 tokens (training) / 128K (inference) |
-| Vocabulary | 128,256 tokens |
-| Precision | BF16 |
-
-### LoRA Adapter Details
-
-| Parameter | Value |
-|-----------|-------|
-| Rank (r) | 128 |
-| Alpha | 256 |
-| Dropout | 0.05 |
-| Target Modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Trainable Parameters | 167,772,160 |
-
-### Files Included
-
-```
-llama-agri-b2b-intelligence/
-├── README.md                 # This file
-├── adapter_config.json       # LoRA configuration
-├── adapter_model.safetensors # LoRA weights
-├── tokenizer_config.json     # Tokenizer settings
-├── special_tokens_map.json   # Special tokens
-└── tokenizer.json            # Full tokenizer
-```
-
----
-
-## Data Sources Referenced in Training
-
-The training data incorporates knowledge from:
+## License
 
-- **USDA NASS** - National Agricultural Statistics Service
-- **USDA ERS** - Economic Research Service
-- **FSA** - Farm Service Agency (subsidies, programs)
-- **RMA** - Risk Management Agency (crop insurance)
-- **SSURGO** - Soil Survey Geographic Database
-- **CDL** - Cropland Data Layer
-- **CME Group** - Commodity futures data
-
----
+This adapter is a Llama 3.1 derivative. It is distributed under the
+[Llama 3.1 Community License](https://www.llama.com/llama3_1/license/), and use must also comply with
+Meta's [Acceptable Use Policy](https://www.llama.com/llama3_1/use-policy/).
+Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
+See also the note on teacher-generated data above.
 
 ## Citation
 
 ```bibtex
-@misc{llama-agri-b2b-intelligence-2024,
-  title={Agricultural B2B Intelligence: Multi-Teacher Knowledge Distillation
-         for Domain-Specific Large Language Models},
-  author={Sarathi Balakrishnan},
-  year={2024},
-  publisher={HuggingFace},
-  url={https://huggingface.co/sarathi-balakrishnan/llama-agri-b2b-intelligence},
-  note={Fine-tuned using Claude Sonnet 4 and GPT-4.1 as teachers on
-        NVIDIA DGX Spark with Blackwell GPU}
+@misc{llama-agri-b2b-intelligence-2025,
+  title  = {Llama 3.1 8B Agricultural B2B Intelligence: a sequence-level distillation experiment},
+  author = {Sarathi Balakrishnan},
+  year   = {2025},
+  publisher = {Hugging Face},
+  url    = {https://huggingface.co/sarathi-balakrishnan/llama-agri-b2b-intelligence}
 }
 ```
-
----
-
-## License
-
-**MIT License** - Free for commercial and research use.
-
-This model is released under the MIT license. The base Llama 3.1 model is subject to Meta's Llama 3.1 Community License Agreement.
-
----
-
-## Acknowledgments
-
-- **Meta AI** - For the Llama 3.1 base model
-- **Anthropic** - For Claude Sonnet 4 teacher responses
-- **OpenAI** - For GPT-4.1 teacher responses
-- **NVIDIA** - For DGX Spark infrastructure
-- **Hugging Face** - For PEFT library and model hosting
-
----
-
-## Contact & Support
-
-- **HuggingFace:** [@sarathi-balakrishnan](https://huggingface.co/sarathi-balakrishnan)
-- **GitHub:** [@sarathi-aiml](https://github.com/sarathi-aiml)
-
----
-
-<div align="center">
-
-**Built with multi-teacher knowledge distillation**
-
-*Transforming frontier AI capabilities into deployable agricultural intelligence*
-
-</div>