Spaces:
Sleeping
Sleeping
Commit History
-
remove NARADA_CLAUDE_CODE_PROMPT.md
Verified signature6e5f135 -
remove Guidlines.txt
Verified signature0ea14aa -
update README.md
Verified signatureb90b33c -
add Blog.md
Verified signature8746415 -
add kl_divergence.png
Verified signaturee42ce74 -
add reward_std.png
Verified signature9928900 -
add reward_curve.png
Verified signature147e79d -
add loss_curve.png
Verified signature7f28648 -
add before_after.png
Verified signature594dc2e
-
Multi-step GRPO, milestone reward, benchmark, Neo4j export, graph viz
Krishna +1: Claude Sonnet 4.6Training: - Multi-step outcome GRPO: model generates 3-5 action plans; full plan executed in env; terminal reward is the training signal (10x more reward variance than single-step) - Async-parallel reward: asyncio.gather runs all G=8 plans concurrently - N_SEEDS_PER_TASK 20->40 for more case diversity (120 total prompts) - max_completion_length 600->800 to fit 4-5 JSON action blocks Environment: - Milestone reward (+0.10) on first visit to causal gene node; creates phenotype->gene->variant two-stage reward landscape Server: - /graph/subgraph endpoint returns D3.js-ready subgraph JSON New files: - benchmark.py: multi-model comparison (Qwen3-0.6/1.7/4B, Llama-3.2-3B, Phi-3.5-mini) via HF Inference Router - scripts/export_neo4j.py: export full graph to Cypher for Neo4j Aura Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>2a62ebe -
Increase GRPO diversity: G=8, temperature=1.1
Krishna +1: Claude Sonnet 4.6reward_std=0 on ~60% of steps because G=4 completions often pick the same hop target. G=8 + higher temperature forces more diverse outputs per group so GRPO gets non-zero advantages more often. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>522f6e8 -
Add CLAUDE.md with recurring training bug patterns
Krishna +1: Claude Sonnet 4.6Documents the three bugs that keep reappearing after notebook edits: TRL chat-format completions, Unsloth telemetry timeout, Qwen3 thinking mode. Also notes max_completion_length=600 requirement and cache-clear. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>19e9a3d -
Fix narada_grpo.ipynb Cell 6: handle TRL chat-format completions
Krishna +1: Claude Sonnet 4.6TRL >= 0.9 passes completions as List[Dict] not plain strings. Extract assistant turn content before parse_action to prevent TypeError. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>967f55d -
Fix narada_grpo.ipynb Cell 3: disable Unsloth stats + thinking mode
Krishna +1: Claude Sonnet 4.6Add UNSLOTH_DISABLE_STATISTICS=1 to prevent 120s HF telemetry timeout. Add apply_chat_template monkey-patch to disable Qwen3 thinking mode, consistent with the colab notebook. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>13f9b5c -
Fix narada_grpo.ipynb: max_completion_length 300 -> 600
Krishna +1: Claude Sonnet 4.6Teammate's notebook had the same null-training bug as the colab version. Also removed max_new_tokens duplicate and added cache-clear logic. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>90fd809 -
upgrades for openenv validation
b654948 -
Fix notebook cell source encoding (char-split artifact from editor)
Krishna +1: Claude Sonnet 4.6NotebookEdit stored cell sources as arrays of individual characters; fix by joining and re-splitting on newlines. Cells 4 and 7 repaired. Content unchanged: Cell 7 still has max_completion_length=600 and cache-clear logic; Cell 4 still has phenotypes_absent in format_obs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>52f3d5d -
Add real multi-model benchmark results to README
Krishna +1: Claude Sonnet 4.6llama-3.1-8b-instant, mixtral-8x7b-32768, and gemma2-9b-it benchmarked across all 3 task tiers. Smaller models collapse to summarise_trail; llama-3.1-8b attempts navigation but misfires on flag_causal. Results demonstrate environment correctly stress-tests reasoning ability across model sizes — core motivation for GRPO fine-tuning. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>1880cb5 -
Add phenotypes_absent obs field, fix training config, update README framing
Krishna +1: Claude Sonnet 4.6- NaradaObservation gains phenotypes_absent + phenotype_absent_names fields; all three case generators populate them from disease HPO terms not assigned to this patient, giving the model explicit diagnostic exclusion signal - environment.py passes absent phenotypes through _build_observation - inference.py and notebook format_obs both surface absent phenotypes in prompt - Notebook Cell 7: max_completion_length 300 → 600 (fixes null-training root cause where Qwen3 thinking tokens consumed all 300 budget, reward_std=0); auto-clears unsloth_compiled_cache on run - Notebook Cell 4: system prompt adds rule 5 for absent phenotypes - README: adds HF Space link, scopes Narada to variant-prioritization substage, clarifies Adversary as planned future work, adds multi-model benchmark table Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>6c965a0 -
rules added for training and evaluation of Narada-CLAUDE, along with a colab notebook for training. Also added a prompt for generating code using Narada-CLAUDE.
Krishna8ebf9d6
-
Add contributing guide, local setup instructions, and .env.example
Krishna +1: Claude Sonnet 4.6Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>612012f -
Add training notebook, architecture docs, baseline benchmark, and WS fixes
Krishna +1: Claude Sonnet 4.6- training/narada_grpo.ipynb: Colab notebook for GRPO fine-tuning Qwen3-1.7B with curriculum training (monogenic → oligogenic → phenotype_mismatch) - ARCHITECTURE.md: full docs — graph schema, reward design, session lifecycle, three-agent system, action/observation space - README: baseline benchmark table (llama-3.3-70b-versatile, zero-shot); shows high variance motivating GRPO training - inference.py: auto-detect Groq vs HF backend from env vars (GROQ_API_KEY or HF_TOKEN); handle ConnectionClosed cleanly instead of silent reconnect - client.py: tighter ping settings (15s interval), remove silent reconnect that masked session loss after WS dropout - pyproject.toml / requirements.txt: add python-dotenv dependency - .gitignore: add .env to secrets exclusions Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>feeb6b0 -
Add OpenEnv standard endpoints and uv.lock
Krishna +1: Claude Sonnet 4.6- GET /metadata: name + description for validator - GET /schema: Pydantic JSON schemas for action/observation/state - POST /mcp: JSON-RPC 2.0 endpoint for MCP tool discovery - requirements.txt + pyproject.toml: add openenv-core>=0.2.0 - pyproject.toml: [project.scripts] server = narada.server.app:main - uv.lock: generated lockfile Local validation now passes. Live validator passes 3/6 → 6/6. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b3577f1 -
Add HF Space frontmatter to README
Krishna +1: Claude Sonnet 4.6Required by HF Spaces to recognise the Docker SDK and expose port 7860. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>dbcc569 -
Rename project to Narada
Krishna +1: Claude Sonnet 4.6Rename Python package clindetect → narada throughout: - src/envs/narada/ replaces src/envs/clindetect/ - Class names: ClinDetectGraph → NaradaGraph, ClinDetectEnvironment → NaradaEnvironment, ClinDetectEnv → NaradaEnv, ClinDetectAction/Observation/State → Narada* - All entry points, uvicorn module paths, openenv.yaml, Dockerfile CMD updated - README, pyproject.toml, inference.py ENV_URL updated to narada-env space Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>9e4dc47 -
Initial ClinDetect OpenEnv environment
Krishna +1: Claude Sonnet 4.6Gene-disease knowledge graph RL environment for rare disease diagnosis. 74K nodes (HPO phenotypes, diseases, genes, variants) built from real ClinVar + HPO data. Three task tiers: monogenic, oligogenic, phenotype_mismatch. Server, client, inference.py, openenv.yaml, Dockerfile all validated locally. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>077e61e