v0.7.0 — Bid V6 IS-DD (match-trained champion)

Bid V6 IS-DD is the new default bid model.

Trained 75M steps (45M + 30M resume) on 5M ISDD-scored deals with:

• score-aware v3 obs (117-dim: scores cumulés + 4 belote bits en main)
• belote-aware reward (Q+K of trump same hand → +20 pts)
• match simulation (cumulative scores + dealer rotation 0→1→2→3 + reset @ 2000)

Arena results (vs V5 IS-DD, post-2026-04-20 dealer-rotation fix)

Eval set	v6 vs v5
DMC play (round-robin)	58.7% / +140
IS-DD play (h2h, 1000 matches)	57.3% / +181
With belief net vs nn_v2_isdd	63.5% / +187

V6 wins arena matches by 13.5pp vs the previous best-overall champion (nn_v2_isdd).

Notes

• Belief net (belief_v3.bin) contributes ~0pp on top of v6 — it added ~13pp on bid_v2. Likely distribution shift since belief was trained on bid_v2 auctions, or v6's score-aware obs already encodes the relevant signal.
• Match-trained models cannot be ranked by deal-level audit metrics — at single-deal probes v6 scores -22 pts/deal vs v5 (audit measures avg deal-EV, not match outcome). Always use match arena for match-trained models.

Assets

• bid_v6_isdd.bin — production bid model (2.4 MB, hidden=512, layers=3, obs_dim=117)

PyPI wheels follow on tag push.

v0.6.0 — Bid V5 IS-DD (score-aware champion)

Default bid model → Bid V5 IS-DD

The default bidder is now Bid V5 IS-DD (25M training steps, score-aware v2 observations, IS-DD-only reward pool).

Key improvements vs v3 Max

Metric	v3 Max	v5 IS-DD
DMC play arena vs v2 champion	−1.5%	+11.0%
IS-DD play arena	+5.9%	+14.6%
Opening XGB-baseline accuracy (h0 probe)	—	99%

What's new

• Score-aware features (5 extra dims: my/opp normalized, win_prob, leader_dist, diff) — obs_dim = 113
• Reward clipping + Polyak EMA (τ = 0.005) + cosine LR decay for training stability
• IS-DD-only reward pool (instead of max(DMC, IS-DD)) — makes the bidder a true IS-DD synergy multiplier

Breaking change

obs_dim changed from 108 → 113. The colver-py bridge now dispatches on net.obs_dim() to build the right observation; old 108-dim models (v2, v3, v4) still work via the same entry point. Web frontend auto-loads v5 by default (match-neutral score context for single-deal play).

Assets

• bid_v5_isdd.bin — Bid V5 IS-DD weights (2.3 MB, 512 hidden × 3 layers, dueling)

Companion docs

• Human-usable strategy guide (88-94% agreement with NN)
• Hidden-layer probe analysis — 2 features discovered that close the XGBoost gap from 77% to 97% on defense
• /annoncer web page — visual strategy reference

v0.5.0 — Bid V3 Max

Default bid model → Bid V3 Max

The default bidder is now Bid V3 Max (20M training steps), replacing Bid V2 (Bid à Dédé).

Trained on max(DMC, IS-DD) real points instead of pure DD values — the only bid model that doesn't lose to v2 in either DMC or IS-DD evaluation. Best synergy with IS-DD play (+5.9%).

Changes

• Model files now use explicit names (bid_v3_max.bin) instead of generic bid_nn_final.bin
• Same architecture as v2 (512 hidden, 3 layers) — drop-in replacement

Assets

• bid_v3_max.bin — Bid V3 Max weights (2.3 MB)

v0.4.0 — Bid à Dédé + DouDou50

New default models

Bid à Dédé — New bidding NN (108→512³→43), trained on 1M DD-solved deals with 24× suit augmentation. 64.5% win rate vs predecessor "Bid à Doudou".

DouDou50 — New play NN (411→1024³→32 ResNet), trained 50M steps with canonical suit encoding. Matches IS-DD solver strength with instant inference (~1ms).

Model files

• bid_nn_final.bin — Bid à Dédé (2.4MB)
• dmc_50.bin — DouDou50 (9.8MB)
• belief_v3.bin — Belief net v3 (1.8MB, unchanged)

Highlights

• BidNet auto-detects hidden size (backward compatible with old models)
• PyO3 supports both legacy (415-dim) and canonical (411-dim) observation formats
• Arena framework for systematic bot evaluation
• Triforge training pipeline for iterative bid+play improvement

v0.3.3 – Belief V3 (3-class)

What's new

• Belief V3: 3-class belief network (left/partner/right) — removes wasted observer class, +count regularization (0.1)
• Trained 213 epochs on 500K games, val_loss=0.8662, val_acc=54.95%
• Auto-detection: loads both V3 (3-class) and legacy (4-class) models transparently
• Model download URLs updated for web frontend

Model assets

• dmc_27.bin — DouDou27 card play model (9.8 MB)
• bid_nn_final.bin — Le Bide à Dédé NN bidder (421 KB)
• belief_v3.bin — Belief V3 card location predictor, 3-class (1.8 MB)

🤖 Generated with Claude Code

Training assets (belief net)

Pre-built training binary + replay data for cloud GPU training

v0.3.1 – DouDou27 (27.5M checkpoint) + Annonces bid eval

What's new

• DouDou27: new DMC card play agent trained for 27.5M steps with better bidding system, see below (replaces DouDou35)
• Annonces tab: bid evaluation now works (NN bid model bundled as release asset)
• Annonces tab hand preview: selected cards shown at full game size below the palette
• Fix bid model discovery to also check ./models/ relative to working directory

Model assets

• dmc_27.bin — DouDou27 card play model (9.8 MB)
• bid_nn_final.bin — Le Bide à Dédé NN bidder (421 KB)

v0.3.0 – DouDou35 (35M checkpoint)

DouDou35 – DMC 35M-step checkpoint

Best checkpoint from tournament evaluation. Replaces the previous dmc_final.bin.

Changes

• Model weights: dmc_35.bin (35M training steps, dueling DQN)
• UI display name: DouDou → DouDou35
• Download URL updated in _model.py and Docker entrypoint

Deployment

Docker containers with COLVER_UPDATE_MODEL=1 will auto-download the new weights on restart.

v0.2.0

Colver v0.2.0

First PyPI-ready release.

Package changes

• Proper Python package structure (pip install colver, pip install colver[web])
• Removed training infra (VecEnv, PrioritizedReplayBuffer) from public API
• colver.download_model() fetches weights from this release
• python -m colver.web starts the web UI
• Type stubs (_colver.pyi) for IDE support
• WTFPL license

Model weights

• dmc_final.bin — DMC Q-network (33M training steps, dueling DQN)
  • Architecture: 415→1024→1024→1024→32 MLP with LayerNorm (~2.6M params)
  • Pure Rust inference (~1ms/decision, no PyTorch needed)
  • Bidding: ImprovedV2 heuristic

Usage

import colver
colver.download_model()  # fetches dmc_final.bin to ~/.cache/colver/models/