60 lines
3.3 KiB
Markdown
60 lines
3.3 KiB
Markdown
# LLM Training Pool Research Expansion
|
|
|
|
Date: 2026-05-10 UTC
|
|
Owner: Codex
|
|
|
|
## Summary
|
|
|
|
Codex expanded MAGATAMA's training pools with curated research input for all five active LLM lanes.
|
|
|
|
The work intentionally avoided copying third-party article/vendor-document bodies. Instead, it added source metadata, provenance-safe source seeds, crawler/evaluation policies, and original SFT examples that teach the desired behavior.
|
|
|
|
## Files Added In Magatama
|
|
|
|
- `training-data/model-registry/external-ingest/llm-lane-research-seeds-2026-05-10.jsonl`
|
|
- `training-data/gitea-learning-pool/magatamallm/curated-web-research-2026-05-10.train.jsonl`
|
|
- `training-data/gitea-learning-pool/magatamallm/curated-web-research-2026-05-10.valid.jsonl`
|
|
- `training-data/gitea-learning-pool/fo_blogllm/curated-web-research-2026-05-10.train.jsonl`
|
|
- `training-data/gitea-learning-pool/fo_blogllm/curated-web-research-2026-05-10.valid.jsonl`
|
|
- `training-data/gitea-learning-pool/tip_llm/curated-web-research-2026-05-10.train.jsonl`
|
|
- `training-data/gitea-learning-pool/tip_llm/curated-web-research-2026-05-10.valid.jsonl`
|
|
- `training-data/gitea-learning-pool/pulso_llm/curated-web-research-2026-05-10.train.jsonl`
|
|
- `training-data/gitea-learning-pool/pulso_llm/curated-web-research-2026-05-10.valid.jsonl`
|
|
- `training-data/gitea-learning-pool/contact_llm/curated-web-research-2026-05-10.train.jsonl`
|
|
- `training-data/gitea-learning-pool/contact_llm/curated-web-research-2026-05-10.valid.jsonl`
|
|
|
|
## Builder Change
|
|
|
|
`scripts/runpod_dataset_builder.ts` now automatically reads supplemental `*.train.jsonl` and `*.valid.jsonl` files from each lane's Gitea learning-pool directory.
|
|
|
|
This means future curated research drops can be added as small lane-specific files without manually editing huge `train.jsonl`/`valid.jsonl` files.
|
|
|
|
## Rebuilt Dataset Counts
|
|
|
|
- `magatamallm`: `1396 train / 156 eval / 1552 total`
|
|
- `fo_blogllm`: `17357 train / 1931 eval / 19288 total`
|
|
- `tip_llm`: `303 train / 35 eval / 338 total`
|
|
- `pulso_llm`: `54 train / 8 eval / 62 total`
|
|
- `contact_llm`: `33 train / 6 eval / 39 total`
|
|
|
|
## Lane Policy Notes
|
|
|
|
- `magatamallm`: cybersecurity, AI security, infrastructure security, proof-before-close, safe remediation, artifact-gated training adoption.
|
|
- `fo_blogllm`: Rene/Flexoptix technical and founder voice, source-backed blogs, no copied source bodies, no fabricated numbers.
|
|
- `tip_llm`: crawler, scraper, parser, robots, search patterns, switch/transceiver/vendor/advisory/forum research.
|
|
- `pulso_llm`: customer-facing Flexoptix/switch/transceiver support, product planning, troubleshooting and quote preparation; never invent SKU, price, stock, warranty or compatibility.
|
|
- `contact_llm`: public business/contact research with source provenance, schema.org/RDAP/PeeringDB/security.txt awareness, robots/privacy guardrails.
|
|
|
|
## Verification
|
|
|
|
- JSONL validation passed for all new files.
|
|
- `pnpm exec tsx scripts/runpod_dataset_builder.ts` passed outside sandbox after `tsx` IPC needed host permissions.
|
|
- `pnpm exec tsx scripts/model_registry_build.ts` passed.
|
|
- `git diff --check` passed.
|
|
|
|
## Follow-Up
|
|
|
|
- Push Magatama training-pool changes to Gitea.
|
|
- Deploy or pull the updated builder and training-data on Erik before starting the next training run.
|
|
- For the next RunPod training run, verify artifact adoption and version bump per strict success rule.
|