fix: replace all 7B references with 1.5B

fix: replace all 7B references with 1.5B

sync: docs, training page fixes, OpenEnv SFT demo notebook

sync: docs, training page fixes, OpenEnv SFT demo notebook

sync: docs, training page fixes, OpenEnv SFT demo notebook

sync: docs, training page fixes, OpenEnv SFT demo notebook

sync: docs, training page fixes, OpenEnv SFT demo notebook

sync: docs, training page fixes, OpenEnv SFT demo notebook

sync: docs, training page fixes, OpenEnv SFT demo notebook

Relocate training notebooks, add BLOG and Google Colab links (SFT + GRPO HF Job), dashboard updates, and eval artifacts

docs: formatting fixes to README.md

feat: made most image results visible

feat: added reward audit program

fix(reward func): reward func was converting data to a lit

fix: add chat template to GRPO prompts

fix: increased max completion length to reduce model output truncation

Add GRPO HF job reward/loss curves, dashboard wiring, plot script, and fix grpo_train log_history unwrap

fix: add pydantic/numpy/fastapi to requirements-train.txt for HF Jobs

feat: added images, new sft notebook, jobs to do grpo

feat: created script to push datasets to huggingface

feat: training results page + SFT Colab notebook

fix: use valid HF emoji in README YAML front matter (◈ → 🤝)

Add OpenEnv client, compat layer, manifest, scripts, GRPO plot hook, and README

docs: fix Nash BATNA formula in parlay_hf_article for GitHub Markdown (HTML, no LaTeX)

fix: reward_fn: robust parse + format gradient; colab T4 GRPO defaults.

fix(grpo): normalize dataset kwargs in reward_fn (TRL may pass 1-elem lists)

fix(grpo): load SFT as base+PEFT so adapter dirs work (no top-level model_type)

fix(grpo): put reward_weights on GRPOConfig when TRL no longer accepts it on GRPOTrainer

fix(grpo): pass num_generations into train_grpo and pin generation_batch_size for TRL G divisibility

fix: changes to grpo_train pipeline to fix divisibility of per_device_train_batch_size and gradient_accumulation_steps

fix(sft): T4/Colab-friendly defaults (grad checkpoint, batch 2/accum 8) + CLI flags

fix(sft): pick SFTConfig max_length vs max_seq_length by TRL version at runtime

docs: minor corrections in the hf article

fix(sft): TRL 1.0+ uses max_length in SFTConfig, not max_seq_length

docs: added badges to README.md

docs: added tags to READMEs

Add pre-training audit scripts, OpenEnv manifest, and tune Parlay training/env (GRPO 1.5B default, min-reward filters, weighted data gen, hiring ZOPA+drift, veteran/opponent prompts, Docker/docs)

feat: added first draft of the hf article

feat: flash-lite for data-gen and flash for UI; remove training page; card tests; --quiet data gen; data/ inspect path; random baseline; GRPO env wrapper; reward fixes (buyer ZOPA, ToM signals); drift + Brier metrics; Bayesian ToM module

fix: trainer notebook improvements

fix: move global declarations before first use (grpo_train, call_gemini)

feat: backup existing data + per-episode progress tracking + gemini live-call verification

fix: normalise reward terms for acquisition_term_sheet scale mismatch

fix: fixed sys.path issues on running generate_data.py

feat: backup pre-2.5 data + add --inspect flag for quality diagnostic run

fix: upgrade gemini model string to 2.5-flash-lite + add tom diagnostic script

fix: gemini retry backoff + tom belief diagnostic logging

fix: resolve WebSocket HTTP 403 on OpenEnv env server

feat: streamline parlay for demo mode and add spectator negotiation mechanics

feat: split Gemini 2.5 Flash (demo) and Flash-Lite (data), SFT threshold 0.3, favicon + check_gemini