feat(training): add SFT demo conversion pipeline

.github/workflows/test.yml

@@ -11,6 +11,8 @@ on:
 jobs:
   test:
     runs-on: ${{ matrix.os }}
+    env:
+      UV_NO_SOURCES: "1"
     strategy:
       matrix:
         os: [ubuntu-latest, macos-latest]
@@ -38,4 +40,4 @@ jobs:
         run: uv run ruff format --check openadapt_ml/
 
       - name: Run pytest
-        run: uv run pytest tests/ -v
+        run: uv run pytest tests/ -v --ignore=tests/integration

.gitignore

@@ -81,3 +81,6 @@ _trash/
 # Auto-generated files
 RESOURCES.md
 azure_logs/
+
+# User-specific screenshot mappings (contains absolute paths)
+screenshot_mapping.json

configs/qwen3vl_demo.yaml

@@ -0,0 +1,32 @@
+model:
+  name: Qwen/Qwen3-VL-2B-Instruct  # 2B for fast iteration; upgrade to 8B once pipeline is validated
+  load_in_4bit: true
+
+lora:
+  r: 16
+  lora_alpha: 32
+  lora_dropout: 0.0
+  bias: none
+  target_modules:
+    - q_proj
+    - v_proj
+    - k_proj
+    - o_proj
+  # task_type is always CAUSAL_LM (set in trl_trainer.py, not configurable)
+  weights_path: checkpoints/qwen3vl2b_demo_lora
+
+training:
+  num_train_epochs: 10  # 20 samples, need more passes
+  per_device_train_batch_size: 1
+  gradient_accumulation_steps: 4
+  learning_rate: 5.0e-5
+  lr_scheduler_type: cosine
+  warmup_ratio: 0.1
+  weight_decay: 0.01
+  max_grad_norm: 0.5
+  logging_steps: 1
+  save_strategy: epoch
+  early_stop_loss: 1.0
+  early_stop_patience: 5
+  early_stop_min_delta: 0.1  # Stop if loss doesn't improve by at least 0.1
+  early_stop_plateau_patience: 5  # for 5 consecutive steps

docs/gpu_hosting_options.md

@@ -0,0 +1,138 @@
+# GPU Compute Hosting Options for OpenAdapt.ai
+
+**Context**: Fine-tuning Qwen3-VL (2B-8B), MIT-licensed open source project, no budget.
+
+**Hardware requirements**:
+- Qwen3-VL 2B with QLoRA: ~8-10 GB VRAM (fits on T4 16GB)
+- Qwen3-VL 8B with QLoRA: ~16-24 GB VRAM (needs A10/A100/L4)
+- Unsloth reduces VRAM by ~60% and speeds training 1.7x
+
+---
+
+## TIER 1: Apply Immediately (Highest Value)
+
+### 1. AWS Cloud Credits for Open Source
+- **Status**: ACTIVE (running since 2019, reaffirmed April 2025)
+- **Credits**: Varies per project; AWS has given millions across 200+ OSS projects
+- **GPUs**: Full AWS fleet (P4d/P5 with A100/H100, G5 with A10G, G6 with L4)
+- **Eligibility**: OSI-approved license (MIT qualifies), active AWS account, valid payment method
+- **Strings attached**: Credits for project infrastructure, not personal use
+- **How to apply**: https://aws.amazon.com/blogs/opensource/aws-promotional-credits-open-source-projects/
+- **Why #1**: Most directly relevant. MIT-licensed, established on GitHub — this program was designed for projects like OpenAdapt.
+
+### 2. Google TPU Research Cloud (TRC)
+- **Status**: ACTIVE (rolling admissions)
+- **Credits**: Free access to 1,000+ Cloud TPU devices
+- **Hardware**: Cloud TPUs (v2, v3, v4) — not GPUs, but excellent for VLM training via JAX/PyTorch XLA
+- **Eligibility**: Anyone can apply; no academic affiliation required
+- **Strings attached**: Must share research publicly (blog posts, open-source code, or papers)
+- **How to apply**: https://sites.research.google/trc/
+- **Why #2**: Completely free, rolling admissions. Sharing publicly is already met by MIT license.
+
+### 3. NVIDIA Inception Program (Unlocks Multi-Platform Credits)
+- **Status**: ACTIVE, free to join
+- **Direct benefits**: Free DLI training credits, SDK access, preferred hardware pricing
+- **Indirect benefits**:
+  - AWS Activate credits: $25,000-$100,000
+  - Nebius AI Lift: up to $150,000 in cloud credits
+  - DGX Cloud Innovation Lab: 2 months of DGX Cloud access (select members)
+- **Eligibility**: At least one developer, working website, officially incorporated, <10 years old
+- **Strings attached**: No equity required
+- **How to apply**: https://www.nvidia.com/en-us/startups/
+- **Why #3**: Force multiplier. Free to join, unlocks $25K-$150K across partners.
+
+### 4. Microsoft for Startups (Azure Credits)
+- **Status**: ACTIVE (restructured July 2025)
+- **Credits**: $1,000 immediately (no application), up to $5,000 with business verification, up to $150,000 for investor-backed startups
+- **GPUs**: Azure NC/ND series (A100, H100, V100, T4)
+- **Eligibility**: No funding required for $5K tier
+- **Strings attached**: Credits expire (90-180 days)
+- **How to apply**: https://www.microsoft.com/en-us/startups (self-service sign-up)
+- **Why #4**: $1K-$5K with no funding needed. Quick to get started.
+
+---
+
+## TIER 2: Strong Options (Apply Soon)
+
+### 5. Google for Startups Cloud Program
+- **Credits**: $2,000 (pre-funding), up to $100,000 (VC-backed), up to $350,000 (AI-first track)
+- **GPUs**: Full GCP fleet (A100, H100, L4, T4, TPU)
+- **Eligibility**: Start tier requires no funding
+- **How to apply**: https://startup.google.com/cloud/
+
+### 6. Lambda Labs Research Grant
+- **Credits**: Up to $5,000 in Cloud Credits plus mentoring
+- **GPUs**: Lambda Cloud fleet (A100, H100, A6000)
+- **Eligibility**: Academic researchers or published research projects; 50% academic discount also available
+- **How to apply**: https://lambda.ai/research
+
+### 7. fal.ai Research Grants
+- **Credits**: Free compute (amount per project)
+- **Eligibility**: Open to anyone — no formal degree required
+- **How to apply**: Email grants@fal.ai with project description. See https://fal.ai/grants
+
+### 8. AMD AI Developer Program & Developer Cloud
+- **Credits**: $100 free (~50 hours) to start; additional credits for public projects
+- **GPUs**: AMD Instinct MI300X (192GB VRAM)
+- **Note**: Check ROCm compatibility with your training stack
+- **How to apply**: https://www.amd.com/en/developer/resources/cloud-access/amd-developer-cloud.html
+
+### 9. a16z Open Source AI Grants
+- **Type**: Cash grants (not investment, not equity)
+- **Eligibility**: Open-source AI developers, hackers, researchers, small teams
+- **How to apply**: Watch https://a16z.com/supporting-the-open-source-ai-community/ for batch announcements
+
+---
+
+## TIER 3: Free Platforms (Use Today, No Application)
+
+| Platform | GPU | Free Limits | Best For |
+|----------|-----|-------------|----------|
+| [Kaggle](https://kaggle.com) | T4/P100 | 30 hrs/week, background exec | **Training Qwen3-VL 2B now** |
+| [Google Colab](https://colab.research.google.com) | T4 | 15-30 hrs/week | Backup training |
+| [Lightning.ai](https://lightning.ai) | T4 | 22 hrs/month | Development |
+| [SageMaker Studio Lab](https://studiolab.sagemaker.aws) | T4 | 4hr sessions, 8hr/day, no credit card | Experimentation |
+| [Modal](https://modal.com) | A100/H100 | $30/month credit (~2hr A100) | Burst training |
+| [Paperspace Gradient](https://paperspace.com/gradient/free-gpu) | M4000 (8GB) | 6hr sessions, restartable | Too small for VLM training |
+
+---
+
+## TIER 4: Research/Academic Programs
+
+| Program | Credits | Eligibility |
+|---------|---------|-------------|
+| [NSF ACCESS](https://access-ci.org) | Large GPU allocations (thousands of hours) | US institution PI |
+| [Amazon Research Awards](https://amazon.science/research-awards) | $70K funds + $50K AWS credits | Academic institutions |
+| [NVIDIA Academic Hardware Grant](https://academicgrants.nvidia.com/academicgrantprogram/s/Application) | Physical GPU hardware | Faculty at PhD-granting institutions |
+| [HOSTKEY GPU Grant](https://hostkey.com/about-us/grants-for-scientific-projects-and-startups/) | Free GPU server time | Research/startup proposals (monthly windows) |
+
+---
+
+## TIER 5: Inference Hosting (Post-Training)
+
+| Platform | Free Tier | Notes |
+|----------|-----------|-------|
+| [HuggingFace ZeroGPU](https://huggingface.co/docs/hub/en/spaces-zerogpu) | H200 inference, daily quota | Best free option for deploying fine-tuned model |
+| [Replicate](https://replicate.com) | Limited free runs | Deploying custom models as APIs |
+| [Fireworks AI](https://fireworks.ai) | 10 RPM free, $1 credit | Supports Qwen models |
+| [Together AI](https://together.ai) | No free tier ($5 min) | Startup accelerator offers up to $50K |
+
+---
+
+## Recommended Action Plan
+
+| Phase | Action | Expected Value | Timeline |
+|-------|--------|---------------|----------|
+| **Today** | Sign up Microsoft for Startups | $1K-$5K Azure | Same day |
+| **Today** | Start training on Kaggle | 30 hrs/week free T4 | Immediate |
+| **This week** | Apply AWS OSS Credits | $5K-$50K+ | 2-4 weeks |
+| **This week** | Apply Google TRC | Free TPU access | 1-2 weeks |
+| **This week** | Email fal.ai (grants@fal.ai) | Free compute | 1-2 weeks |
+| **This month** | Join NVIDIA Inception | Unlocks $25K-$150K partner credits | 2-4 weeks |
+| **This month** | Apply Google for Startups | $2K-$350K | 1-2 weeks |
+| **This month** | Apply Lambda Research Grant | $5K | 2-4 weeks |
+| **Post-training** | Deploy on HuggingFace ZeroGPU | Free H200 inference | When model ready |
+
+---
+
+*Research conducted February 2026. Verify program availability before applying.*

docs/training_pipeline_gaps.md

@@ -0,0 +1,90 @@
+# Training Pipeline Gaps & Plan
+
+## Current State (Feb 23, 2026)
+
+First successful JSONL training launch on Lambda Labs (Qwen3-VL-8B, A10).
+Pipeline works end-to-end: `convert_demos --bundle` -> `train.py --jsonl` -> Lambda launch.
+
+### What Works
+- Bundle creation (JSONL + images with relative paths)
+- JSONL loader in TRL trainer
+- Lambda Labs instance launch, code sync, bundle upload
+- Checkpoint save (TRL SFTTrainer auto-saves per epoch + final)
+- Dashboard with Lambda cloud badge, cost display, setup progress
+- Checkpoint download via rsync
+
+### What's Missing
+
+## Gap 1: TRL Training Callback (logging + early stopping)
+
+**Problem**: `train_from_jsonl()` uses TRL's SFTTrainer which logs to stdout/TensorBoard
+but NOT to our `training_log.json`. The Lambda polling loop in `lambda_labs.py` reads
+`training_log.json` for dashboard updates — so the dashboard shows 0 progress.
+
+Also, `early_stop_loss` is in YAML config but not wired to a TRL callback.
+
+**Fix**: Add a `TrainerCallback` subclass to `trl_trainer.py`:
+- `on_log()`: Write step/epoch/loss/lr to `training_log.json`
+- `on_log()`: Check `early_stop_loss` threshold, set `trainer.state.should_training_stop = True`
+- ~40 lines
+
+**Files**: `openadapt_ml/training/trl_trainer.py`
+
+## Gap 2: Cost Persistence
+
+**Problem**: Cost is only computed in dashboard JavaScript. Not saved to `training_log.json`.
+After instance terminates, cost is lost.
+
+**Fix**: Write `cost_per_hour`, `total_cost`, `instance_type` to `training_log.json` in the
+TRL callback (same one from Gap 1). The callback can read instance_type from an env var
+or config field.
+
+**Files**: `openadapt_ml/training/trl_trainer.py` (extend callback)
+
+## Gap 3: Post-Training Eval Automation
+
+**Problem**: After training completes on Lambda, there's no automatic evaluation. User must
+manually run `eval_policy.py` or `compare.py`.
+
+**Fix**: Add `--eval` flag to `train.py` that runs eval after training completes:
+1. Load the saved checkpoint
+2. Run prediction on each training sample (sanity check: can it reproduce training data?)
+3. Write results to `eval_results.json` alongside checkpoint
+4. ~30 lines in `train.py`, reusing `eval_policy.py` logic
+
+For Lambda: add eval step to `run_training()` command string when `--eval` is passed.
+
+**Files**: `openadapt_ml/scripts/train.py`, `openadapt_ml/cloud/lambda_labs.py`
+
+## Gap 4: Training Speed (Unsloth on Lambda)
+
+**Problem**: 8B model without Unsloth = 569s/step on A10. With Unsloth should be ~100-200s/step.
+Current Lambda setup installs deps with `uv sync` which doesn't include Unsloth (it's optional
+and needs special install: `pip install unsloth`).
+
+**Fix**: Add Unsloth install to Lambda setup script. Remove `--no-unsloth` from default
+Lambda train command. Test Unsloth + Qwen3-VL-8B on A10.
+
+Alternative: Use 2B model (`Qwen/Qwen3-VL-2B-Instruct`) which is 4x faster without Unsloth.
+
+**Files**: `openadapt_ml/cloud/lambda_labs.py` (setup_instance), `configs/qwen3vl_demo.yaml`
+
+## Gap 5: Auto-Terminate After Training
+
+**Problem**: When training is launched manually (SSH nohup), the Lambda instance keeps running
+after training completes. No auto-terminate.
+
+**Fix**: The automated `train` command already handles this when it's used end-to-end.
+The issue is only when we SSH in manually. Not a code gap — operational.
+
+## Priority Order
+
+1. **Gap 4** (speed) — blocking: current run takes 8 hours, wastes money
+2. **Gap 1** (callback) — high: dashboard shows nothing during training
+3. **Gap 3** (eval) — medium: need eval to know if training worked
+4. **Gap 2** (cost) — low: nice-to-have for tracking spend
+5. **Gap 5** (terminate) — low: operational
+
+## Immediate Action
+
+Kill current slow run, switch to 2B model, relaunch. Fix Gaps 1+4 first.