Upload README.md

Update blog.md

Upload blog.md

docs: correct HF endpoint deployment command

docs: add quick HF CLI endpoint deployment steps

publish: mark merged models as transformers text-generation artifacts

add provider interface

publish: inline chat template into tokenizer_config for HF chat compatibility

Add w and b report link

docs: replace in-progress training notes with final run results

remove the redundant section in the dashboard

Add qwen3-8b red vs blue showdown output

docs: refine wording in README for clarity and flow

docs: humanize README origin line

docs: remove Fibr mention and correct Round 1 origin story

docs: remove misplaced tagline from README intro

docs: replace Dario reference with direct security-intuition thesis

docs: add a few plan quips without turning README into a joke

docs: tighten README back to direct no-nonsense style

docs: rewrite README narrative to feel human, fix Blue curriculum framing

docs: restore Mythos as Claude's model name in origin story

docs: clarify dual-role env design, name mesh components explicitly, fix Claude typo

docs: clarify Dario Amodei reference and correct model name to Claude

docs: replace WarGames README with Faultline story-first README

docs: add training.md operator quickstart, remove superpowers planning artifacts

config: max_steps 100 -> 60 to fit budget at empirical 150s/step

env_client: retry transient 5xx + typed EnvUnavailableError; trainer survives env outages

action_parser: strict variant strips Qwen3 <think>...</think> before json.loads

trainer: A+B+C+D — soften think-mask, strict parser, add format reward, max_steps 100

smoke: drop vllm_gpu_memory_utilization 0.50->0.35 (a10g headroom for 1024 ctx)

rollout: mask think block from GRPO loss + free GPU before merge subprocess

rollout: disable Qwen3 thinking mode in chat template (clipped_ratio=0.75 -> expected drop)

publish: in-job adapter+merged push via subprocess; upload_large_folder; env overrides

publish: auto push adapter (and optional merged) to HF Hub after training

tsconfig fix

fix the tsc issue

fix cf pages issue of the tsc, use npx instead

Update tsconfig.tsbuildinfo

jobs: wait for CUDA init before trainer; trainer: add per-component reward funcs

base: max_steps 200 -> 150 for budget headroom

base: tighten for 5h budget — max_steps=200, completion=768, ep_steps=4

smoke: grad_accum=4 so generation_batch divisible by num_generations

smoke: resize for a10g — Qwen3-1.7B, max_completion=512, ep_steps=3

smoke: mirror base config shape with Qwen3-1.7B, 5 steps

base: max_completion_length=1024, max_seq=10240, ep_steps=5, gc=true

Add tiny smoke yaml: Qwen3-0.6B on a10g

smoke: cap vllm_max_model_length=8192, util=0.55, gc=true; passthrough both

smoke: enable gradient_checkpointing, lower vllm gpu util to 0.30, ep_steps=2

Flatten rollout logprobs to 2D list[list[float]] for TRL sampling_per_token_logps

Fix GRPO rollout: vLLM multi-turn sampling, env_mask, peft_config to TRL, max_completion_length=500