Metadata-Version: 2.4
Name: erl-trainer
Version: 0.1.2
Summary: Experiential Reinforcement Learning (ERL) — a thin wrapper on HuggingFace TRL's GRPOTrainer implementing the ERL algorithm with reflection, memory, and internalization.
Project-URL: Repository, https://github.com/akarim23131/erl_trainer
Author: Abid Karim
License: MIT License
        
        Copyright (c) 2025 Abid Karim
        
        Permission is hereby granted, free of charge, to any person obtaining a copy
        of this software and associated documentation files (the "Software"), to deal
        in the Software without restriction, including without limitation the rights
        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
        copies of the Software, and to permit persons to whom the Software is
        furnished to do so, subject to the following conditions:
        
        The above copyright notice and this permission notice shall be included in all
        copies or substantial portions of the Software.
        
        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
        SOFTWARE.
License-File: LICENSE
Keywords: erl,grpo,language-models,reflection,reinforcement-learning,trl
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.10
Requires-Dist: datasets
Requires-Dist: peft
Requires-Dist: torch>=2.0.0
Requires-Dist: transformers>=4.50.0
Requires-Dist: trl>=0.15.0
Description-Content-Type: text/markdown

# erl-trainer

**Experiential Reinforcement Learning (ERL)** — a thin wrapper on HuggingFace TRL's `GRPOTrainer` that adds a reflection-retry-internalization loop to standard GRPO training.

Install it, swap `GRPOTrainer` for `ERLTrainer`, add a `feedback_func`, and you get ERL training. Everything else — LoRA, quantization, datasets, reward functions — works exactly like TRL.

## Installation

```bash
pip install erl-trainer
```

## Quick Start

```python
from erl import ERLConfig, ERLTrainer
from datasets import load_dataset
from peft import LoraConfig  # optional

dataset = load_dataset("your_dataset", split="train")

# Standard reward function (same as TRL)
def reward_func(prompts, completions, **kwargs):
    return [compute_your_score(c) for c in completions]

# NEW: Textual feedback function (unique to ERL)
def feedback_func(prompts, completions, **kwargs):
    return [get_your_feedback(c) for c in completions]

config = ERLConfig(
    output_dir="erl-output",
    num_train_epochs=3,
    learning_rate=1e-6,
    per_device_train_batch_size=4,
    num_generations=4,
    # ERL-specific params
    reward_threshold=1.0,
    memory_size=50,
    memory_top_k=3,
    internalization_coef=1.0,
)

# Optional: LoRA config (works exactly like TRL)
lora_config = LoraConfig(
    r=64,
    lora_alpha=64,
    target_modules="all-linear",
)

trainer = ERLTrainer(
    model="Qwen/Qwen2.5-3B-Instruct",
    args=config,
    train_dataset=dataset,
    reward_funcs=reward_func,
    feedback_func=feedback_func,   # NEW: the only addition vs TRL
    peft_config=lora_config,       # optional, same as TRL
)

trainer.train()
```

## The ERL Algorithm

Each training step runs seven phases:

| Phase | Description |
|-------|-------------|
| **1. First attempt** | Generate responses `y1` for the batch; compute numerical reward `r1` and textual feedback `f1`. These are two separate signals — `r1` drives the RL math, `f1` explains *why* the attempt failed. |
| **2. Gating** | Samples where `r1 < reward_threshold` enter the reflection loop; others are done. No wasted compute on already-successful attempts. |
| **3. Self-reflection** | For gated samples, the model reflects using all five inputs: the original prompt, `y1`, `f1`, `r1`, and relevant entries retrieved from cross-episode memory. Produces a natural-language improvement plan `Δ`. |
| **4. Second attempt** | The model generates `y2` conditioned on the original prompt and `Δ` only (not `y1` or `f1`). Reward `r2` is computed against the original task. |
| **5. Memory update** | If `r2 > threshold`, the reflection `Δ` is stored in a FIFO cross-episode memory. Future steps retrieve the most recent stored reflections to seed the reflection prompt. |
| **6. GRPO update** | Policy gradient over the combined batch — `y1` (reward `r1`), `Δ` (reward `r2`), and `y2` (reward `r2`) — in one joint update. Negative advantage pushes the model away from bad outputs; positive advantage reinforces good ones. |
| **7. Internalization** | SFT cross-entropy on `(original_prompt → y2)` pairs for successful second attempts (`r2 > 0`). Trains the model to produce the improved answer directly from `x`, without any reflection scaffold at inference time. |

Reflections and retries are generated by the **same model with the same weights** as the first attempt — no freezing, no separate model. All generation happens before the optimizer step.

## Configuration

All `GRPOConfig` options are inherited. ERL adds:

| Parameter | Default | Description |
|-----------|---------|-------------|
| `reward_threshold` | `1.0` | Gating threshold τ. Samples with `r1 >= τ` skip reflection. |
| `memory_size` | `50` | Max reflections stored in cross-episode memory. |
| `memory_top_k` | `3` | Reflections retrieved per reflection prompt. |
| `reflection_system_prompt` | *(built-in)* | Template with `{prompt}`, `{attempt}`, `{feedback}`, `{reward}`, `{memory}`. |
| `retry_system_prompt` | *(built-in)* | Template with `{prompt}` and `{reflection}`. |
| `internalization_coef` | `1.0` | Weight of internalization loss relative to RL loss. |
| `enable_memory` | `True` | Toggle cross-episode memory on/off. |
| `enable_internalization` | `True` | Toggle the distillation step on/off. |

## Feedback Function

The only new concept vs TRL is `feedback_func`. It receives the same keyword arguments as a reward function and must return a list of strings — one per completion:

```python
def feedback_func(prompts, completions, **kwargs) -> list[str]:
    feedbacks = []
    for prompt, completion in zip(prompts, completions):
        feedbacks.append(f"Your answer was missing: {diagnose(prompt, completion)}")
    return feedbacks
```

`feedback_func` may be `None`. In that case, empty strings are passed to the reflection prompt — training still runs, but reflection quality will be lower since the model has no textual diagnosis to work from.

## Implementation Notes

### TRL version compatibility

The trainer accesses several TRL-internal methods (`_generate`, `_calculate_rewards`, `_get_per_token_logps_and_entropies`). Compatibility guards are in place for the most likely breaking changes:

- `_generate` return values are accessed by index rather than by position destructuring, so additions to TRL's return tuple are silently ignored.
- `_get_per_token_logps_and_entropies` output is handled whether TRL returns a `(logps, entropies)` tuple or a dict.
- The input batch is normalised at entry so both `dict[str, list]` (standard PyTorch DataLoader collate) and `list[dict]` (TRL's collator) are accepted.

Tested against `trl>=0.15.0`. Pin your TRL version in production.

### Batched reflection and retry generation

Reflections and retries for all gated samples in a batch are generated in two batched `model.generate` calls (one for all reflections, one for all retries), not one call per sample. This keeps GPU utilisation high regardless of how many samples are gated.

### Algorithm 1 vs Algorithm 2

This implementation follows **Algorithm 1** (simplified) from the paper: `y1`, `Δ`, and `y2` are packed into one combined batch and updated in a single GRPO pass with global advantage normalisation.

The paper's Appendix A describes **Algorithm 2** (full), which runs two separate RL updates — one on `y1` alone, then one on `Δ + y2` — giving per-group advantage normalisation for each update. The directional gradients are the same; only the advantage scale differs. Algorithm 2 can be implemented by splitting the packed batch and calling `compute_loss` twice.

### Ablations

Both memory and internalization can be disabled independently, which is useful for ablation studies:

```python
config = ERLConfig(
    enable_memory=False,          # no cross-episode memory
    enable_internalization=False, # no distillation step, pure RL
    ...
)
```

## License

MIT
